🕷️ Crawler Inspector

URL Lookup

Direct Parameter Lookup

Raw Queries and Responses

1. Shard Calculation

Query:

Response:

Calculated Shard: 169 (from laksa062)

2. Crawled Status Check

Query:

curl -X POST \
  'http://laksa169.int.ahrefs:8124/' \
  -H 'Content-Type: text/plain' \
  -H 'X-ClickHouse-Database: crawler3' \
  -H 'Authorization: Basic YXBpOg==' \
  -d 'SELECT getAhrefsURLFromUnparsed(src_unparsed) AS found_url, ifNull(toUnixTimestamp(download_stamp), 0) AS crawl_time, ifNull(toUnixTimestamp(props_url_first_seen), 0) AS first_indexed_time, download_http_code AS http_code, src_unparsed AS src_unparsed, src_root_hash AS src_root_hash, history_drop_reason AS history_drop_reason, meta_title AS meta_title, meta_descriptions AS meta_descriptions, attrs_boilerpipe_text AS attrs_boilerpipe_text, attrs_markdown AS attrs_markdown, attrs_readable_markdown AS attrs_readable_markdown, meta_canonical AS meta_canonical FROM crawler3.page_info_local FINAL PREWHERE (src_root_hash, src_unparsed) IN ((getAhrefsRootHashFromUnparsed(getAhrefsUnparsedNoserviceFromURL(\'https://catboost.ai/docs/en/references/training-parameters/common\')), getAhrefsUnparsedNoserviceFromURL(\'https://catboost.ai/docs/en/references/training-parameters/common\'))) FORMAT JSONEachRow'

Response:

{"found_url":"https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/common","crawl_time":1776291628,"first_indexed_time":1731949884,"http_code":200,"src_unparsed":"ai,catboost!\/docs\/en\/references\/training-parameters\/common s443","src_root_hash":"17435841955170310369","history_drop_reason":null,"meta_title":"Common parameters | CatBoost","meta_descriptions":["loss_function. Command-line: --loss-function. Alias: objective. Description."],"attrs_boilerpipe_text":"loss_function\nCommand-line:\n--loss-function\nAlias:\nobjective\nDescription\nThe\nmetric\nto use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the \nObjectives and metrics\nsection for details on each metric).\nFormat:\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\nSupported metrics\nRMSE\nLogloss\nMAE\nCrossEntropy\nQuantile\nLogLinQuantile\nLq\nMultiRMSE\nMultiClass\nMultiClassOneVsAll\nMultiLogloss\nMultiCrossEntropy\nMAPE\nPoisson\nPairLogit\nPairLogitPairwise\nQueryRMSE\nQuerySoftMax\nGroupQuantile\nTweedie\nYetiRank\nYetiRankPairwise\nStochasticFilter\nStochasticRank\nA custom python object can also be set as the value of this parameter (see an \nexample\n).\nFor example, use the following construction to calculate the value of Quantile with the coefficient \nα\n=\n0.1\n\\alpha = 0.1\n:\nQuantile:alpha=0.1\nType\nstring\nobject\nDefault value\nPython package\nDepends on the class:\nCatBoostClassifier\n: Logloss if the \ntarget_border\nparameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass.\nCatBoost\nand\nCatBoostRegressor\n: RMSE\nR package, Command-line\nRMSE\nSupported processing units\nCPU and GPU\ncustom_metric\nCommand-line:\n--custom-metric\nDescription\nMetric\nvalues to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the \nObjectives and metrics\nsection for details on each metric).\nFormat:\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\nSupported metrics\nExamples\nCalculate the value of CrossEntropy:\nCrossEntropy\nCalculate the value of Quantile with the coefficient \nα\n=\n0.1\n\\alpha = 0.1\nQuantile:alpha=0.1\nCalculate the values of Logloss and AUC:\n[\n'Logloss'\n,\n'AUC'\n]\nValues of all custom metrics for learn and validation datasets are saved to the \nMetric\noutput files (\nlearn_error.tsv\nand\ntest_error.tsv\nrespectively). The directory for these files is specified in the \n--train-dir\n(\ntrain_dir\n) parameter.\nUse the \nvisualization tools\nto see a live chart with the dynamics of the specified metrics.\nType\nstring\nlist of strings\nDefault value\nPython package\nNone\nR package\nNone\nCommand-line\nNone (do not output additional metric values)\nSupported processing units\nCPU and GPU\neval_metric\nCommand-line:\n--eval-metric\nDescription\nThe metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the \nObjectives and metrics\nsection for details on each metric).\nFormat:\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\nSupported metrics\nA user-defined function can also be set as the value (see an \nexample\n).\nExamples:\nR2\nType\nstring\nobject\nDefault value\nOptimized objective is used\nSupported processing units\nCPU and GPU\niterations\nCommand-line:\n-i\n,\n--iterations\nAliases:\nnum_boost_round\n,\nn_estimators\n,\nnum_trees\nDescription\nThe maximum number of trees that can be built when solving machine learning problems.\nWhen using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.\nType\nint\nDefault value\n1000\nSupported processing units\nCPU and GPU\nlearning_rate\nCommand-line:\n-w\n,\n--learning-rate\nAlias:\neta\nDescription\nThe learning rate.\nUsed for reducing the gradient step.\nType\nfloat\nDefault value\nThe default value is defined automatically for\nLogloss\n,\nMultiClass\nand\nRMSE\nloss functions depending on the number of iterations if none of parameters\nleaf_estimation_iterations\n,\nleaf_estimation_method\n,\nl2_leaf_reg\nis set. In this case, the selected learning rate is printed to stdout and saved in the model.\nIn other cases, the default value is 0.03.\nSupported processing units\nCPU and GPU\nrandom_seed\nCommand-line:\n-r\n,\n--random-seed\nAlias:\nrandom_state\nDescription\nThe random seed used for training.\nType\nint\nDefault value\nPython package\nNone (0)\nR package, Command-line\n0\nSupported processing units\nCPU and GPU\nl2_leaf_reg\nCommand-line:\n--l2-leaf-reg\n,\nl2-leaf-regularizer\nAlias:\nreg_lambda\nDescription\nCoefficient at the L2 regularization term of the cost function.\nAny positive value is allowed.\nType\nfloat\nDefault value\n3.0\nSupported processing units\nCPU and GPU\nbootstrap_type\nCommand-line:\n--bootstrap-type\nDescription\nBootstrap type\n. Defines the method for sampling the weights of objects.\nSupported methods:\nBayesian\nBernoulli\nMVS\nPoisson (supported for GPU only)\nNo\nType\nstring\nDefault value\nThe default value depends on\nobjective\n,\ntask_type\n,\nbagging_temperature\nand\nsampling_unit\n:\nWhen the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5.\nNeither MultiClass nor MultiClassOneVsAll, task_type = CPU and sampling_unit = Object: MVS with the subsample parameter set to 0.8.\nOtherwise: Bayesian.\nSupported processing units\nCPU and GPU\nbagging_temperature\nCommand-line:\n--bagging-temperature\nDescription\nDefines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes.\nUse the Bayesian bootstrap to assign random weights to objects.\nThe weights are sampled from exponential distribution if the value of this parameter is set to\n1\n. All weights are equal to 1 if the value of this parameter is set to\n0\n.\nPossible values are in the range\n[\n0\n;\ninf\n⁡\n)\n[0; \\inf)\n. The higher the value the more aggressive the bagging is.\nThis parameter can be used if the selected bootstrap type is Bayesian.\nType\nfloat\nDefault value\n1\nSupported processing units\nCPU and GPU\nsubsample\nCommand-line:\n--subsample\nDescription\nSample rate for bagging.\nThis parameter can be used if one of the following bootstrap types is selected:\nPoisson\nBernoulli\nMVS\nType\nfloat\nDefault value\nThe default value depends on the dataset size and the bootstrap type:\nDatasets with less than 100 objects — 1\nDatasets with 100 objects or more:\nPoisson, Bernoulli — 0.66\nMVS — 0.8\nSupported processing units\nCPU and GPU\nsampling_frequency\nCommand-line:\n--sampling-frequency\nDescription\nFrequency to sample weights and objects when building trees.\nSupported values:\nPerTree — Before constructing each new tree\nPerTreeLevel — Before choosing each new split of a tree\nType\nstring\nDefault value\nPerTreeLevel\nSupported processing units\nCPU\nsampling_unit\nCommand-line:\n--sampling-unit\nDescription\nThe sampling scheme.\nPossible values:\nObject — The weight \nw\ni\nw_{i}\nof the i-th object\no\ni\no_{i}\nis used for sampling the corresponding object.\nGroup — The weight\nw\nj\nw_{j}\nof the group\ng\nj\ng_{j}\nis used for sampling each object \no\ni\nj\no_{i_{j}}\nfrom the group \ng\nj\ng_{j}\n.\nType\nString\nDefault value\nObject\nSupported processing units\nCPU and GPU\nmvs_reg\nCommand-line:\n--mvs-reg\nDescription\nAffects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to\n∞\n\\infty\n- Bernoulli).\nNote\nThis parameter is supported only for the MVS sampling method (the\nbootstrap_type\nparameter must be set to MVS).\nType\nfloat\nDefault value\nThe value is set based on the gradient distribution on the current iteration\nSupported processing units\nCPU\nrandom_strength\nCommand-line:\n--random-strength\nDescription\nThe amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model.\nThe value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected.\nThe scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance.\nNote\nThis parameter is not supported for the following loss functions:\nQueryCrossEntropy\nYetiRankPairwise\nPairLogitPairwise\nType\nfloat\nDefault value\n1\nSupported processing units\nCPU\nuse_best_model\nCommand-line:\n--use-best-model\nDescription\nIf this parameter is set, the number of trees that are saved in the resulting model is defined as follows:\nBuild the number of trees defined by the training parameters.\nUse the validation dataset to identify the iteration with the optimal value of the metric specified in  \n--eval-metric\n(\n--eval-metric\n).\nNo trees are saved after this iteration.\nThis option requires a validation dataset to be provided.\nType\nbool\nDefault value\nTrue if a validation set is input (the eval_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise.\nSupported processing units\nCPU and GPU\nbest_model_min_trees\nCommand-line:\n--best-model-min-trees\nDescription\nThe minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees.\nShould be used with the\n--use-best-model\nparameter.\nType\nint\nDefault value\nPython package, R package\nNone (The minimal number of trees for the best model is not set)\nCommand-line\nThe minimal number of trees for the best model is not set\nSupported processing units\nCPU and GPU\ndepth\nCommand-line:\n-n\n,\n--depth\nAlias:\nmax_depth\nDescription\nDepth of the trees.\nThe range of supported values depends on the processing unit type and the type of the selected loss function:\nCPU — Any integer up to  16.\nGPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions.\nType\nint\nDefault value\n6 (16 if the growing policy is set to Lossguide)\nSupported processing units\nCPU and GPU\ngrow_policy\nCommand-line:\n--grow-policy\nDescription\nThe tree growing policy. Defines how to perform greedy tree construction.\nPossible values:\nSymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric.\nDepthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement.\nNote\nModels with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\nLossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split.\nNote\nModels with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\nType\nstring\nDefault value\nSymmetricTree\nSupported processing units\nCPU and GPU\nmin_data_in_leaf\nCommand-line:\n--min-data-in-leaf\nAlias:\nmin_child_samples\nDescription\nThe minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value.\nCan be used only with the Lossguide and Depthwise growing policies.\nType\nint\nDefault value\n1\nSupported processing units\nCPU and GPU\nmax_leaves\nCommand-line:\n--max-leaves\nAlias:\nnum_leaves\nDescription\nThe maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy.\nNote\nIt is not recommended to use values greater than 64, since it can significantly slow down the training process.\nType\nint\nDefault value\n31\nSupported processing units\nCPU and GPU\nignored_features\nCommand-line:\n-I\n,\n--ignore-features\nDescription\nFeature indices to exclude from the training.\nPython package\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers.  Otherwise, it is assumed that all passed values are feature indices.\nSpecifics:\nNon-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \n42\n, the corresponding non-existing feature is successfully ignored.\nThe identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to\nfeatureCount – 1\n. If a file is used as \ninput data\nthen any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order:\ncat feature<\\t>label value<\\t>num feature\n. So for the row\nrock<\\t>0<\\t>42\n, the identifier for the\nrock\nfeature is 0, and for the\n42\nfeature it's 1.\nFor example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored:\n[1,2,7,42,43,44,45]\nR package\nSpecifics:\nNon-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \n42\n, the corresponding non-existing feature is successfully ignored.\nThe identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to\nfeatureCount – 1\n. If a file is used as \ninput data\nthen any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order:\ncat feature<\\t>label value<\\t>num feature\n. So for the row\nrock<\\t>0<\\t>42\n, the identifier for the\nrock\nfeature is 0, and for the\n42\nfeature it's 1.\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45).\nCommand-line\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers.  Otherwise, it is assumed that all passed values are feature indices.\nSpecifics:\nNon-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \n42\n, the corresponding non-existing feature is successfully ignored.\nThe identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to\nfeatureCount – 1\n. If a file is used as \ninput data\nthen any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order:\ncat feature<\\t>label value<\\t>num feature\n. So for the row\nrock<\\t>0<\\t>42\n, the identifier for the\nrock\nfeature is 0, and for the\n42\nfeature it's 1.\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction:\n1:2:7:42-45\n.\nDefault value\nPython package, R package\nNone\nCommand-line\nOmitted\nSupported processing units\nCPU and GPU\none_hot_max_size\nCommand-line:\n--one-hot-max-size\nDescription\nUse one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features.\nSee\ndetails\n.\nType\nint\nDefault value\nThe default value depends on various conditions:\nN\/A if training is performed on CPU in Pairwise scoring mode\nRead more about  Pairwise scoring\nThe following loss functions use Pairwise scoring:\nYetiRankPairwise\nPairLogitPairwise\nQueryCrossEntropy\nPairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions.\n255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training\n10 if training is performed in \nRanking\nmode\n2 if none of the conditions above is met\nSupported processing units\nCPU and GPU\nhas_time\nCommand-line:\n--has-time\nDescription\nUse the order of objects in the input data (do not perform random permutations during the\nTransforming categorical features to numerical features\nand\nChoosing the tree structure\nstages).\nThe Timestamp column type is used to determine the order of objects if specified in the \ninput data\n.\nType\nbool\nDefault value\nFalse (not used; generates random permutations)\nSupported processing units\nCPU and GPU\nrsm\nCommand-line:\n--rsm\nAlias:\ncolsample_bylevel\nDescription\nRandom subspace method. The percentage of features to use at each split selection, when features are selected over again at random.\nThe value must be in the range (0;1].\nType\nfloat (0;1]\nDefault value\nNone (set to 1)\nSupported processing units\nCPU; GPU for pairwise ranking\nnan_mode\nCommand-line:\n--nan-mode\nDescription\nThe method for \nprocessing missing values\nin the input dataset.\nPossible values:\n\"Forbidden\" — Missing values are not supported, their presence is interpreted as an error.\n\"Min\" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\n\"Max\" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\nUsing the  Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree.\nType\nstring\nDefault value\nMin\nSupported processing units\nCPU and GPU\ninput_borders\nCommand-line:\n--input-borders-file\nDescription\nLoad \nCustom quantization borders and missing value modes\nfrom a file (do not generate them).\nBorders are automatically generated before training if this parameter is not set.\nType\nstring\nDefault value\nPython package\nNone\nCommand-line\nThe file is not loaded, the values are generated\nSupported processing units\nCPU and GPU\noutput_borders\nCommand-line:\n--output-borders-file\nDescription\nSave quantization borders for the current dataset to a file.\nRefer to the\nfile format description\n.\nType\nstring\nDefault value\nPython package\nNone\nCommand-line\nThe file is not saved\nSupported processing units\nCPU and GPU\nfold_permutation_block\nCommand-line:\n--fold-permutation-block\nDescription\nObjects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation.\nType\nint\nDefault value\nPython package\n1\nR package, Command-line\nDefault value differs depending on the dataset size and ranges from 1 to 256 inclusively\nSupported processing units\nCPU and GPU\nleaf_estimation_method\nCommand-line:\n--leaf-estimation-method\nDescription\nThe method used to calculate the values in leaves.\nPossible values:\nNewton\nGradient\nExact\nType\nstring\nDefault value\nDepends on the mode and the selected loss function:\nRegression with Quantile or MAE loss functions — One Exact iteration.\nRegression with any loss function but Quantile or MAE – One Gradient iteration.\nClassification mode – Ten Newton iterations.\nMulticlassification mode – One Newton iteration.\nSupported processing units\nCPU and GPU\nleaf_estimation_iterations\nCommand-line:\n--leaf-estimation-iterations\nDescription\nCatBoost might calculate leaf values using several gradient or newton steps instead of a single one.\nThis parameter regulates how many steps are done in every tree when calculating leaf values.\nType\nint\nDefault value\nPython package\nNone (Depends on the training objective)\nR package, Command-line\nDepends on the training objective\nSupported processing units\nCPU and GPU\nleaf_estimation_backtracking\nCommand-line:\n--leaf-estimation-backtracking\nDescription\nWhen the value of the\nleaf_estimation_iterations\nparameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree.\nThe behaviour differs depending on the value of this parameter:\nNo — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf.\nAny other value —Backtracking is used.\nIn this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf.\nWhen\nleaf_estimation_iterations\nfor the Command-line version is set to\nn\n, the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having\nn\ngradient steps, the algorithm makes a single gradient step that is reduced\nn\ntimes, which means that it is divided by\n2\n⋅\nn\n2\\cdot n\ntimes.\nPossible values:\nNo — Do not use backtracking. Supported on CPU and GPU.\nAnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU.\nArmijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU.\nType\nstring\nDefault value\nAnyImprovement\nSupported processing units\nDepends on the selected value\nfold_len_multiplier\nCommand-line:\n--fold-len-multiplier\nDescription\nCoefficient for changing the length of folds.\nThe value must be greater than 1. The best validation result is achieved with minimum values.\nWith values close to 1 (for example,\n1\n+\nϵ\n1+\\epsilon\n), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects.\nType\nfloat\nDefault value\n2\nSupported processing units\nCPU and GPU\napprox_on_full_history\nCommand-line:\n--approx-on-full-history\nDescription\nThe principles for calculating the approximated values.\nPossible values:\nFalse\n — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows:\n1\nX\n\\frac{1}X\n, where \nX\nis the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate\nTrue\n — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate.\nType\nbool\nDefault value\nPython package, Command-line\nFalse\nR package\nTrue\nSupported processing units\nCPU\nclass_weights\nCommand-line:\n--class-weights\nDescription\nClass weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems.\nPython package\nNote\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to\n(\ns\nu\nm\n_\nn\ne\ng\na\nt\ni\nv\ne\ns\nu\nm\n_\np\no\ns\ni\nt\ni\nv\ne\n)\n\\left(\\frac{sum\\_negative}{sum\\_positive}\\right)\nfor class 1.\nFor example,\nclass_weights=[0.1, 4]\nmultiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\nIf class labels are not standard consecutive integers [0, 1 ... class_count-1], use the dict or collections.OrderedDict type with label to weight mapping.\nFor example,\nclass_weights={'a': 1.0, 'b': 0.5, 'c': 2.0}\nmultiplies the weights of objects with class label\na\nby 1.0, the weights of objects with class label\nb\nby 0.5 and the weights of objects with class label\nc\nby 2.0.\nThe dictionary form can also be used with standard consecutive integers class labels for additional readability. For example:\nclass_weights={0: 1.0, 1: 0.5, 2: 2.0}\n.\nNote\nClass labels are extracted from dictionary keys for the following types of class_weights:\ndict\ncollections.OrderedDict (when the order of classes in the model is important)\nThe class_names parameter can be skipped when using these types.\nAlert\nDo not use this parameter with auto_class_weights and scale_pos_weight.\nR package\nFor example,\nclass_weights <- c(0.1, 4)\nmultiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\nAlert\nDo not use this parameter with auto_class_weights.\nCommand-line\nNote\nThe quantity of class weights must match the quantity of class names specified in the\n--class-names\nparameter and the number of classes specified in the\n--classes-count parameter\n.\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to\n(\ns\nu\nm\n_\nn\ne\ng\na\nt\ni\nv\ne\ns\nu\nm\n_\np\no\ns\ni\nt\ni\nv\ne\n)\n\\left(\\frac{sum\\_negative}{sum\\_positive}\\right)\nfor class 1.\nFormat:\n<value for class 1>,..,<values for class N>\nFor example:\n0.85,1.2,1\nAlert\nDo not use this parameter with auto_class_weights.\nType\nlist\ndict\ncollections.OrderedDict\nDefault value\nNone (the weight for all classes is set to 1)\nSupported processing units\nCPU and GPU\nclass_names\nDescription\nClasses names. Allows to redefine the default values when using the MultiClass and Logloss metrics.\nIf the upper limit for the numeric class label is specified, the number of classes names should match this value.\nWarning\nThe quantity of classes names must match the quantity of classes weights specified in the \n--class-weights\nparameter and the number of classes specified in the \n--classes-count\nparameter.\nFormat:\n<name for class 1>,..,<name for class N>\nFor example:\nsmartphone,touchphone,tablet\nType\nlist of strings\nDefault value\nNone\nSupported processing units\nCPU and GPU\nauto_class_weights\nCommand-line:\n--auto-class-weights\nDescription\nAutomatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights.\nSupported values:\nNone — All class weights are set to 1\nBalanced:\nC\nW\nk\n=\nm\na\nx\nc\n=\n1\nK\n(\n∑\nt\ni\n=\nc\nw\ni\n)\n∑\nt\ni\n=\nk\nw\ni\nCW_k=\\displaystyle\\frac{max_{c=1}^K(\\sum_{t_{i}=c}{w_i})}{\\sum_{t_{i}=k}{w_{i}}}\nSqrtBalanced:\nC\nW\nk\n=\nm\na\nx\nc\n=\n1\nK\n(\n∑\nt\ni\n=\nc\nw\ni\n)\n∑\nt\ni\n=\nk\nw\ni\nCW_k=\\sqrt{\\displaystyle\\frac{max_{c=1}^K(\\sum_{t_i=c}{w_i})}{\\sum_{t_i=k}{w_i}}}\nAlert\nDo not use this parameter with \nclass_weights\nand\nscale_pos_weight\n.\nType\nstring\nDefault value\nNone — All class weights are set to 1\nSupported processing units\nCPU and GPU\nscale_pos_weight\nDescription\nThe weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1.\nNote\nFor imbalanced datasets, the weight multiplier can be set to \n(\ns\nu\nm\n_\nn\ne\ng\na\nt\ni\nv\ne\ns\nu\nm\n_\np\no\ns\ni\nt\ni\nv\ne\n)\n\\left(\\frac{sum\\_negative}{sum\\_positive}\\right)\nAlert\nDo not use this parameter with\nauto_class_weights\nand\nclass_weights\n.\nType\nfloat\nDefault value\n1.0\nSupported processing units\nCPU and GPU\nboosting_type\nCommand-line:\n--boosting-type\nDescription\nBoosting scheme.\nPossible values:\nOrdered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme.\nPlain — The classic gradient boosting scheme.\nType\nstring\nDefault value\nDepends on the processing unit type, the number of objects in the training dataset and the selected learning mode\nCPU\nPlain\nGPU\nAny number of objects, MultiClass or MultiClassOneVsAll mode: Plain\nMore than 50 thousand objects, any mode: Plain\nLess than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered\nSupported processing units\nCPU and GPU\nOnly the Plain mode is supported for the MultiClass loss on GPU\nboost_from_average\nCommand-line:\n--boost-from-average\nDescription\nInitialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value.\nAvailable for the following loss functions:\nRMSE\nLogloss\nCrossEntropy\nQuantile\nMAE\nMAPE\nType\nbool\nDefault value\nDepends on the selected loss function:\nTrue for RMSE, Quantile, MAE, MAPE\nFalse for all other loss functions\nSupported processing units\nCPU and GPU\nlangevin\nCommand-line:\n--langevin\nDescription\nEnables the Stochastic Gradient Langevin Boosting mode.\nRefer to the\nSGLB: Stochastic Gradient Langevin Boosting\npaper for details.\nType\nbool\nDefault value\nFalse\nSupported processing units\nCPU\ndiffusion_temperature\nCommand-line:\n--diffusion-temperature\nDescription\nThe diffusion temperature of the Stochastic Gradient Langevin Boosting mode.\nOnly non-negative values are supported.\nType\nfloat\nDefault value\n10000\nSupported processing units\nCPU\nposterior_sampling\nCommand-line:\n--posterior-sampling\nDescription\nIf this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties.\nSpecifies options:\nLangevin\n: true,\nDiffusionTemperature\n: objects in learn pool count,\nModelShrinkRate\n: 1 \/ (2. * objects in learn pool count).\nType\nbool\nDefault value\nFalse\nSupported processing units\nCPU only\nallow_const_label\nCommand-line:\n--allow-const-label\nDescription\nUse it to train models with datasets that have equal label values for all objects.\nType\nbool\nDefault value\nFalse\nSupported processing units\nCPU and GPU\nscore_function\nCommand-line:\n--score-function\nDescription\nThe\nscore type\nused to select the next split during the tree construction.\nPossible values:\nCosine (do not use this score type with the Lossguide tree growing policy)\nL2\nNewtonCosine (do not use this score type with the Lossguide tree growing policy)\nNewtonL2\nType\nstring\nDefault value\nCosine\nSupported processing units\nThe supported score functions vary depending on the processing unit type:\nGPU — All score types\nCPU — Cosine, L2\nmonotone_constraints\nCommand-line:\n--monotone-constraints\nDescription\nImpose monotonic constraints on numerical features.\nPossible values:\n1\n — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features.\n-1\n — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features.\n0\n — constraints are disabled.\nSupported formats for setting the value of this parameter (all feature indices are zero-based):\nSet constraints individually for each feature as a string (the number of features is n).\nFormat\n\"(<constraint_0>, <constraint_2>, .., <constraint_n-1>)\"\nZero constraints for features at the end of the list may be dropped.\nIn\nmonotone_constraints = \"(1,0,-1)\"\nan increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features.\nSet constraints individually for each explicitly specified feature as a string (the number of features is n).\n\"<feature index or name>:<constraint>, .., <feature index or name>:<constraint>\"\nThese examples\nmonotone-constraints = \"2:1,4:-1\"\nmonotone-constraints = \"Feature2:1,Feature4:-1\"\nare identical, given that the name of the feature index 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nSet constraints individually for each required feature as an array or a dictionary (the number of features is n).\nFormat\n[<constraint_0>, <constraint_2>, .., <constraint_n-1>]\n{\"<feature index or name>\":<constraint>, .., \"<feature index or name>\":<constraint>}\nArray examples\nmonotone_constraints = [1, 0, -1]\nThese dictionary examples\nmonotone_constraints = {\n\"Feature2\"\n:\n1\n,\n\"Feature4\"\n:-\n1\n}\nmonotone_constraints = {\n\"2\"\n:\n1\n,\n\"4\"\n:-\n1\n}\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nType\nlist of strings\nstring\ndict\nlist\nDefault value\nPython package, R package\nNone\nCommand-line\nOmmited\nSupported processing units\nCPU\nfeature_weights\nCommand-line:\n--feature-weights\nDescription\nPer-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split.\nNon-negative float values are supported for each weight.\nSupported formats for setting the value of this parameter:\nSet the multiplication weight for each feature as a string (the number of features is n).\nFormat\n\"(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)\"\nNote\nSpaces between values are not allowed.\nValues should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped.\nIn this\nexample\nfeature_weights = \"(0.1,1,3)\"\nthe multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\nSet the multiplication weight individually for each explicitly specified feature as a string (the number of features is n).\nFormat\n\"<feature index or name>:<weight>, .., <feature index or name>:<weight>\"\nNote\nSpaces between values are not allowed.\nThese examples\nfeature_weights = \"2:1.1,4:0.1\"\nfeature_weights = \"Feature2:1.1,Feature4:0.1\"\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nSet the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n).\nFormat\n[<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>]\n{\"<feature index or name>\":<weight>, .., \"<feature index or name>\":<weight>}\nArray examples\nfeature_weights = [0.1, 1, 3]\nThese dictionary examples\nfeature_weights = {\n\"Feature2\"\n:\n1.1\n,\n\"Feature4\"\n:\n0.3\n}\nfeature_weights = {\n\"2\"\n:\n1.1\n,\n\"4\"\n:\n0.3\n}\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nType\nlist\nnumpy.ndarray\nstring\ndict\nDefault value\n1 for all features\nSupported processing units\nCPU\nfirst_feature_use_penalties\nCommand-line:\n--first-feature-use-penalties\nDescription\nPer-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model.\nRefer to the\nPer-object and per-feature penalties\nsection for details on applying different score penalties.\nNon-negative float values are supported for each penalty.\nSet the penalty for each feature as a string (the number of features is n).\nFormat\n\"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\nNote\nSpaces between values are not allowed.\nValues should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\nIn this example\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"(0.1,1,3)\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"(0.1,1,3)\"\nNote\nSpaces between values are not allowed.\nthe multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\nSet the penalty individually for each explicitly specified feature as a string (the number of features is n).\nFormat\n\"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\nNote\nSpaces between values are not allowed.\nThese examples\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"2:1.1,4:0.1\"\nfirst_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"2:1.1,4:0.1\"\nper_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nSet the penalty individually for each required feature as an array or a dictionary (the number of features is n).\nFormat\n[<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n{\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\nArray examples.\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = [0.1, 1, 3]\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = [0.1, 1, 3]\nThese dictionary examples\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = {\n\"Feature2\"\n:\n1.1\n,\n\"Feature4\"\n:\n0.1\n}\nfirst_feature_use_penalties = {\n\"2\"\n:\n1.1\n,\n\"4\"\n:\n0.1\n}\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = {\n\"Feature2\"\n:\n1.1\n,\n\"Feature4\"\n:\n0.1\n}\nper_object_feature_penalties = {\n\"2\"\n:\n1.1\n,\n\"4\"\n:\n0.1\n}\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nType\nlist\nnumpy.ndarray\nstring\ndict\nDefault value\n0 for all features\nSupported processing units\nCPU\nfixed_binary_splits\nCommand-line:\n--fixed-binary-splits\nDescription\nA list of indices of binary features to put at the top of each tree; ignored if\ngrow_policy\nis\nSymmetric\n.\nType\nlist\nDefault value\nNone\nSupported processing units\nGPU\npenalties_coefficient\nCommand-line:\n--penalties-coefficient\nDescription\nA single-value common coefficient to multiply all penalties.\nNon-negative values are supported.\nType\nfloat\nDefault value\n1\nSupported processing units\nCPU\nper_object_feature_penalties\nCommand-line:\n--per-object-feature-penalties\nDescription\nPer-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time.\nRefer to the\nPer-object and per-feature penalties\nsection for details on applying different score penalties.\nNon-negative float values are supported for each penalty.\nPython package\nSet the penalty for each feature as a string (the number of features is n).\nFormat\n\"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\nNote\nSpaces between values are not allowed.\nValues should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\nIn this example\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"(0.1,1,3)\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"(0.1,1,3)\"\nNote\nSpaces between values are not allowed.\nthe multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\nSet the penalty individually for each explicitly specified feature as a string (the number of features is n).\nFormat\n\"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\nNote\nSpaces between values are not allowed.\nThese examples\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"2:1.1,4:0.1\"\nfirst_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"2:1.1,4:0.1\"\nper_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nSet the penalty individually for each required feature as an array or a dictionary (the number of features is n).\nFormat\n[<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n{\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\nArray examples.\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = [0.1, 1, 3]\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = [0.1, 1, 3]\nThese dictionary examples\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = {\n\"Feature2\"\n:\n1.1\n,\n\"Feature4\"\n:\n0.1\n}\nfirst_feature_use_penalties = {\n\"2\"\n:\n1.1\n,\n\"4\"\n:\n0.1\n}\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = {\n\"Feature2\"\n:\n1.1\n,\n\"Feature4\"\n:\n0.1\n}\nper_object_feature_penalties = {\n\"2\"\n:\n1.1\n,\n\"4\"\n:\n0.1\n}\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nR package\nSet the penalty for each feature as a string (the number of features is n).\nFormat\n\"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\nNote\nSpaces between values are not allowed.\nValues should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\nPenalties equal to 0 at the end of the list may be dropped.\nIn this\nexample\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"(0.1,1,3)\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"(0.1,1,3)\"\nNote\nSpaces between values are not allowed.\nthe multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\nSet the penalty individually for each explicitly specified feature as a string (the number of features is n).\nFormat\n\"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\nNote\nSpaces between values are not allowed.\nThese examples\nfirst_feature_use_penalties\nparameter:\nfirst_feature_use_penalties = \"2:1.1,4:0.1\"\nfirst_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\nper_object_feature_penalties\nparameter:\nper_object_feature_penalties = \"2:1.1,4:0.1\"\nper_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\nare identical, given that the name of the feature indexed 2 is\nFeature2\nand the name of the feature indexed 4 is\nFeature4\n.\nType\nlist\nnumpy.ndarray\nstring\ndict\nDefault value\n0 for all objects\nSupported processing units\nCPU\nmodel_shrink_rate\nCommand-line:\n--model-shrink-rate\nDescription\nThe constant used to calculate the coefficient for multiplying the model on each iteration.\nThe actual model shrinkage coefficient calculated at each iteration depends on the value of the\n--model-shrink-mode\nfor the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1].\nType\nfloat\nDefault value\nThe default value depends on the values of the following parameters:\n--model-shrink-mode\nfor the Command-line version\n--monotone-constraints\nfor the Command-line version\nSupported processing units\nCPU\nmodel_shrink_mode\nCommand-line:\nmodel_shrink_mode\nDescription\nDetermines how the actual model shrinkage coefficient is calculated at each iteration.\nPossible values:\nConstant:\n1\n−\nm\no\nd\ne\nl\n_\ns\nh\nr\ni\nn\nk\n_\nr\na\nt\ne\n⋅\nl\ne\na\nr\nn\ni\nn\ng\n_\nr\na\nt\ne\n,\n1 - model\\_shrink\\_rate \\cdot learning\\_rate {,}\nm\no\nd\ne\nl\n_\ns\nh\nr\ni\nn\nk\n_\nr\na\nt\ne\nmodel\\_shrink\\_rate\nis the value of the\n--model-shrink-rate\nfor the Command-line version parameter.\nl\ne\na\nr\nn\ni\nn\ng\n_\nr\na\nt\ne\nlearning\\_rate\nis the value of the\n--learning-rate\nfor the Command-line version parameter\nDecreasing:\n1\n−\nm\no\nd\ne\nl\n_\ns\nh\nr\ni\nn\nk\n_\nr\na\nt\ne\ni\n,\n1 - \\frac{model\\_shrink\\_rate}{i} {,}\nm\no\nd\ne\nl\n_\ns\nh\nr\ni\nn\nk\n_\nr\na\nt\ne\nmodel\\_shrink\\_rate\nis the value of the\n--model-shrink-rate\nfor the Command-line version parameter.\ni\ni\nis the identifier of the iteration.\nType\nstring\nDefault value\nConstant\nSupported processing units\nCPU","attrs_markdown":"[![Logo icon](https:\/\/yastatic.net\/s3\/locdoc\/daas-static\/catboost\/71b237a322eec6f2889af0dae2a9c549.svg)](https:\/\/catboost.ai\/ \"CatBoost\")\n\n- Installation\n  - [Overview](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/installation)\n  - Python package installation\n  - CatBoost for Apache Spark installation\n  - R package installation\n  - Command-line version binary\n  - Build from source\n- Key Features\n- Training parameters\n  - [Overview](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/)\n  - [Common parameters](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common)\n  - [CTR settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/ctr)\n  - [Input settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/input)\n  - [Multiclassification settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/multiclassification)\n  - [Output settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/output)\n  - [Overfitting detection settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/overfitting-detection)\n  - [Performance settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/performance)\n  - [Processing unit settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/processing-unit)\n  - [Quantization settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/quantization)\n  - [Text processing parameters](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/text-processing)\n  - [Visualization settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/visualization)\n- Python package\n- CatBoost for Apache Spark\n- R package\n- Command-line version\n- Applying models\n- Objectives and metrics\n- Model analysis\n- Data format description\n- [Parameter tuning](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/parameter-tuning)\n- [Speeding up the training](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/speed-up-training)\n- Data visualization\n- Algorithm details\n- [FAQ](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/faq)\n- Educational materials\n- [Development and contributions](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/development-and-contributions)\n- [Contacts](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/contacts)\n\nloss\\_function\n\n## In this article:\n- [loss\\_function](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#loss_function)\n- [custom\\_metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#custom_metric)\n- [eval\\_metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#eval_metric)\n- [iterations](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#iterations)\n- [learning\\_rate](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#learning_rate)\n- [random\\_seed](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#random_seed)\n- [l2\\_leaf\\_reg](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#l2_leaf_reg)\n- [bootstrap\\_type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#bootstrap_type)\n- [bagging\\_temperature](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#bagging_temperature)\n- [subsample](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#subsample)\n- [sampling\\_frequency](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#sampling_frequency)\n- [sampling\\_unit](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#sampling_unit)\n- [mvs\\_reg](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#mvs_reg)\n- [random\\_strength](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#random_strength)\n- [use\\_best\\_model](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#use_best_model)\n- [best\\_model\\_min\\_trees](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#best_model_min_trees)\n- [depth](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#depth)\n- [grow\\_policy](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#grow_policy)\n- [min\\_data\\_in\\_leaf](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#min_data_in_leaf)\n- [max\\_leaves](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#max_leaves)\n- [ignored\\_features](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#ignored_features)\n- [one\\_hot\\_max\\_size](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#one_hot_max_size)\n- [has\\_time](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#has_time)\n- [rsm](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#rsm)\n- [nan\\_mode](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#nan_mode)\n- [input\\_borders](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#input_borders)\n- [output\\_borders](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#output_borders)\n- [fold\\_permutation\\_block](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fold_permutation_block)\n- [leaf\\_estimation\\_method](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_method)\n- [leaf\\_estimation\\_iterations](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_iterations)\n- [leaf\\_estimation\\_backtracking](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_backtracking)\n- [fold\\_len\\_multiplier](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fold_len_multiplier)\n- [approx\\_on\\_full\\_history](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#approx_on_full_history)\n- [class\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#class_weights)\n- [class\\_names](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#class_names)\n- [auto\\_class\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#auto_class_weights)\n- [scale\\_pos\\_weight](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#scale_pos_weight)\n- [boosting\\_type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#boosting_type)\n- [boost\\_from\\_average](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#boost_from_average)\n- [langevin](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#langevin)\n- [diffusion\\_temperature](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#diffusion_temperature)\n- [posterior\\_sampling](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#posterior_sampling)\n- [allow\\_const\\_label](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#allow_const_label)\n- [score\\_function](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#score_function)\n- [monotone\\_constraints](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#monotone_constraints)\n- [feature\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#feature_weights)\n- [first\\_feature\\_use\\_penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#first_feature_use_penalties)\n- [fixed\\_binary\\_splits](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fixed_binary_splits)\n- [penalties\\_coefficient](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#penalties_coefficient)\n- [per\\_object\\_feature\\_penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#per_object_feature_penalties)\n- [model\\_shrink\\_rate](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#model_shrink_rate)\n- [model\\_shrink\\_mode](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#model_shrink_mode)\n\n1. [Training parameters](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/)\n2. Common parameters\n\n# Common parameters\n- [loss\\_function](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#loss_function)\n- [custom\\_metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#custom_metric)\n- [eval\\_metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#eval_metric)\n- [iterations](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#iterations)\n- [learning\\_rate](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#learning_rate)\n- [random\\_seed](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#random_seed)\n- [l2\\_leaf\\_reg](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#l2_leaf_reg)\n- [bootstrap\\_type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#bootstrap_type)\n- [bagging\\_temperature](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#bagging_temperature)\n- [subsample](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#subsample)\n- [sampling\\_frequency](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#sampling_frequency)\n- [sampling\\_unit](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#sampling_unit)\n- [mvs\\_reg](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#mvs_reg)\n- [random\\_strength](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#random_strength)\n- [use\\_best\\_model](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#use_best_model)\n- [best\\_model\\_min\\_trees](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#best_model_min_trees)\n- [depth](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#depth)\n- [grow\\_policy](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#grow_policy)\n- [min\\_data\\_in\\_leaf](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#min_data_in_leaf)\n- [max\\_leaves](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#max_leaves)\n- [ignored\\_features](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#ignored_features)\n- [one\\_hot\\_max\\_size](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#one_hot_max_size)\n- [has\\_time](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#has_time)\n- [rsm](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#rsm)\n- [nan\\_mode](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#nan_mode)\n- [input\\_borders](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#input_borders)\n- [output\\_borders](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#output_borders)\n- [fold\\_permutation\\_block](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fold_permutation_block)\n- [leaf\\_estimation\\_method](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_method)\n- [leaf\\_estimation\\_iterations](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_iterations)\n- [leaf\\_estimation\\_backtracking](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_backtracking)\n- [fold\\_len\\_multiplier](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fold_len_multiplier)\n- [approx\\_on\\_full\\_history](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#approx_on_full_history)\n- [class\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#class_weights)\n- [class\\_names](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#class_names)\n- [auto\\_class\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#auto_class_weights)\n- [scale\\_pos\\_weight](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#scale_pos_weight)\n- [boosting\\_type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#boosting_type)\n- [boost\\_from\\_average](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#boost_from_average)\n- [langevin](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#langevin)\n- [diffusion\\_temperature](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#diffusion_temperature)\n- [posterior\\_sampling](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#posterior_sampling)\n- [allow\\_const\\_label](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#allow_const_label)\n- [score\\_function](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#score_function)\n- [monotone\\_constraints](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#monotone_constraints)\n- [feature\\_weights](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#feature_weights)\n- [first\\_feature\\_use\\_penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#first_feature_use_penalties)\n- [fixed\\_binary\\_splits](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#fixed_binary_splits)\n- [penalties\\_coefficient](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#penalties_coefficient)\n- [per\\_object\\_feature\\_penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#per_object_feature_penalties)\n- [model\\_shrink\\_rate](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#model_shrink_rate)\n- [model\\_shrink\\_mode](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#model_shrink_mode)\n\n## loss\\_function\nCommand-line: `--loss-function`\n\n*Alias:* `objective`\n\n#### Description\nThe [metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions) to use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\nSupported metrics\n\n- RMSE\n- Logloss\n- MAE\n- CrossEntropy\n- Quantile\n- LogLinQuantile\n- Lq\n- MultiRMSE\n- MultiClass\n- MultiClassOneVsAll\n- MultiLogloss\n- MultiCrossEntropy\n- MAPE\n- Poisson\n- PairLogit\n- PairLogitPairwise\n- QueryRMSE\n- QuerySoftMax\n- GroupQuantile\n- Tweedie\n- YetiRank\n- YetiRankPairwise\n- StochasticFilter\n- StochasticRank\n\nA custom python object can also be set as the value of this parameter (see an [example](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/python-usages-examples)).\n\nFor example, use the following construction to calculate the value of Quantile with the coefficient α \\= 0\\.1 \\\\alpha = 0.1 α\\=0\\.1:\n```\nQuantile:alpha=0.1\n```\n**Type**\n\n- string\n- object\n\n**Default value**\n\nPython package\n\nDepends on the class:\n\n- [CatBoostClassifier](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/python-reference_catboostclassifier): Logloss if the `target_border` parameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass.\n- [CatBoost](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/python-reference_catboost) and [CatBoostRegressor](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/python-reference_catboostregressor): RMSE\n\nR package, Command-line\n\nRMSE\n\n**Supported processing units**\n\nCPU and GPU\n\n## custom\\_metric\nCommand-line: `--custom-metric`\n\n#### Description\n[Metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions) values to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\n[Supported metrics](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/custom-metric__supported-metrics)\n\nExamples\n\n- Calculate the value of CrossEntropy:\n  ```\n  CrossEntropy\n  ```\n- Calculate the value of Quantile with the coefficient α \\= 0\\.1 \\\\alpha = 0.1 α\\=0\\.1\n  ```\n  Quantile:alpha=0.1\n  ```\n- Calculate the values of Logloss and AUC:\n  ```\n  ['Logloss', 'AUC']\n  ```\n\nValues of all custom metrics for learn and validation datasets are saved to the [Metric](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/output-data_loss-function) output files (`learn_error.tsv` and `test_error.tsv` respectively). The directory for these files is specified in the `--train-dir` (`train_dir`) parameter.\n\nUse the [visualization tools](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/features\/visualization) to see a live chart with the dynamics of the specified metrics.\n\n**Type**\n\n- string\n- list of strings\n\n**Default value**\n\nPython package\n\nNone\n\nR package\n\nNone\n\nCommand-line\n\nNone (do not output additional metric values)\n\n**Supported processing units**\n\nCPU and GPU\n\n## eval\\_metric\nCommand-line: `--eval-metric`\n\n#### Description\nThe metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\n[Supported metrics](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/eval-metric__supported-metrics)\n\nA user-defined function can also be set as the value (see an [example](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/python-usages-examples)).\n\nExamples:\n```\nR2\n```\n**Type**\n\n- string\n- object\n\n**Default value**\n\nOptimized objective is used\n\n**Supported processing units**\n\nCPU and GPU\n\n## iterations\nCommand-line: `-i`, `--iterations`\n\n*Aliases:* `num_boost_round`, `n_estimators`, `num_trees`\n\n#### Description\nThe maximum number of trees that can be built when solving machine learning problems.\n\nWhen using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.\n\n**Type**\n\nint\n\n**Default value**\n\n1000\n\n**Supported processing units**\n\nCPU and GPU\n\n## learning\\_rate\nCommand-line: `-w`, `--learning-rate`\n\n*Alias:* `eta`\n\n#### Description\nThe learning rate.\n\nUsed for reducing the gradient step.\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value is defined automatically for [`Logloss`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions-classification#Logit), [`MultiClass`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions-multiclassification#MultiClass) and [`RMSE`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions-regression#RMSE) loss functions depending on the number of iterations if none of parameters [`leaf_estimation_iterations`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_iterations), [`leaf_estimation_method`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#leaf_estimation_method), [`l2_leaf_reg`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/common#l2_leaf_reg) is set. In this case, the selected learning rate is printed to stdout and saved in the model.\n\nIn other cases, the default value is 0.03.\n\n**Supported processing units**\n\nCPU and GPU\n\n## random\\_seed\nCommand-line: `-r`, `--random-seed`\n\n*Alias:*`random_state`\n\n#### Description\nThe random seed used for training.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\nNone (0)\n\nR package, Command-line\n\n0\n\n**Supported processing units**\n\nCPU and GPU\n\n## l2\\_leaf\\_reg\nCommand-line: `--l2-leaf-reg`, `l2-leaf-regularizer`\n\n*Alias:* `reg_lambda`\n\n#### Description\nCoefficient at the L2 regularization term of the cost function.\n\nAny positive value is allowed.\n\n**Type**\n\nfloat\n\n**Default value**\n\n3\\.0\n\n**Supported processing units**\n\nCPU and GPU\n\n## bootstrap\\_type\nCommand-line: `--bootstrap-type`\n\n#### Description\n[Bootstrap type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-main-stages_bootstrap-options). Defines the method for sampling the weights of objects.\n\nSupported methods:\n\n- Bayesian\n- Bernoulli\n- MVS\n- Poisson (supported for GPU only)\n- No\n\n**Type**\n\nstring\n\n**Default value**\n\nThe default value depends on `objective`, `task_type`, `bagging_temperature` and `sampling_unit`:\n\n- When the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging\\_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5.\n- Neither MultiClass nor MultiClassOneVsAll, task\\_type = CPU and sampling\\_unit = Object: MVS with the subsample parameter set to 0.8.\n- Otherwise: Bayesian.\n\n**Supported processing units**\n\nCPU and GPU\n\n## bagging\\_temperature\nCommand-line: `--bagging-temperature`\n\n#### Description\nDefines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes.\n\nUse the Bayesian bootstrap to assign random weights to objects.\n\nThe weights are sampled from exponential distribution if the value of this parameter is set to \"1\". All weights are equal to 1 if the value of this parameter is set to \"0\".\n\nPossible values are in the range \\[ 0 ; inf ⁡ ) \\[0; \\\\inf) \\[0;inf). The higher the value the more aggressive the bagging is.\n\nThis parameter can be used if the selected bootstrap type is Bayesian.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU and GPU\n\n## subsample\nCommand-line: `--subsample`\n\n#### Description\nSample rate for bagging.\n\nThis parameter can be used if one of the following bootstrap types is selected:\n\n- Poisson\n- Bernoulli\n- MVS\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value depends on the dataset size and the bootstrap type:\n\n- Datasets with less than 100 objects — 1\n- Datasets with 100 objects or more:\n  - Poisson, Bernoulli — 0.66\n  - MVS — 0.8\n\n**Supported processing units**\n\nCPU and GPU\n\n## sampling\\_frequency\nCommand-line: `--sampling-frequency`\n\n#### Description\nFrequency to sample weights and objects when building trees.\n\nSupported values:\n\n- PerTree — Before constructing each new tree\n- PerTreeLevel — Before choosing each new split of a tree\n\n**Type**\n\nstring\n\n**Default value**\n\nPerTreeLevel\n\n**Supported processing units**\n\nCPU\n\n## sampling\\_unit\nCommand-line: `--sampling-unit`\n\n#### Description\nThe sampling scheme.\n\nPossible values:\n\n- Object — The weight\n  w\n  \n  i\n  \n  w\\_{i}\n  \n  wi\n  of the i-th object\n  o\n  \n  i\n  \n  o\\_{i}\n  \n  oi\n  is used for sampling the corresponding object.\n- Group — The weight\n  w\n  \n  j\n  \n  w\\_{j}\n  \n  wj\n  of the group\n  g\n  \n  j\n  \n  g\\_{j}\n  \n  gj\n  is used for sampling each object\n  o\n  \n  i\n  \n  j\n  \n  o\\_{i\\_{j}}\n  \n  oij\n  from the group\n  g\n  \n  j\n  \n  g\\_{j}\n  \n  gj\n  .\n\n**Type**\n\nString\n\n**Default value**\n\nObject\n\n**Supported processing units**\n\nCPU and GPU\n\n## mvs\\_reg\nCommand-line: `--mvs-reg`\n\n#### Description\nAffects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to ∞ \\\\infty ∞ - Bernoulli).\n\nNote\n\nThis parameter is supported only for the MVS sampling method (the `bootstrap_type` parameter must be set to MVS).\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe value is set based on the gradient distribution on the current iteration\n\n**Supported processing units**\n\nCPU\n\n## random\\_strength\nCommand-line: `--random-strength`\n\n#### Description\nThe amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model.\n\nThe value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected.\n\nThe scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance.\n\nNote\n\nThis parameter is not supported for the following loss functions:\n\n- QueryCrossEntropy\n- YetiRankPairwise\n- PairLogitPairwise\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU\n\n## use\\_best\\_model\nCommand-line: `--use-best-model`\n\n#### Description\nIf this parameter is set, the number of trees that are saved in the resulting model is defined as follows:\n\n1. Build the number of trees defined by the training parameters.\n2. Use the validation dataset to identify the iteration with the optimal value of the metric specified in `--eval-metric` (`--eval-metric`).\n\nNo trees are saved after this iteration.\n\nThis option requires a validation dataset to be provided.\n\n**Type**\n\nbool\n\n**Default value**\n\nTrue if a validation set is input (the eval\\_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise.\n\n**Supported processing units**\n\nCPU and GPU\n\n## best\\_model\\_min\\_trees\nCommand-line: `--best-model-min-trees`\n\n#### Description\nThe minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees.\n\nShould be used with the `--use-best-model` parameter.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package, R package\n\nNone (The minimal number of trees for the best model is not set)\n\nCommand-line\n\nThe minimal number of trees for the best model is not set\n\n**Supported processing units**\n\nCPU and GPU\n\n## depth\nCommand-line: `-n`, `--depth`\n\n*Alias:* `max_depth`\n\n#### Description\nDepth of the trees.\n\nThe range of supported values depends on the processing unit type and the type of the selected loss function:\n\n- CPU — Any integer up to 16.\n- GPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions.\n\n**Type**\n\nint\n\n**Default value**\n\n6 (16 if the growing policy is set to Lossguide)\n\n**Supported processing units**\n\nCPU and GPU\n\n## grow\\_policy\nCommand-line: `--grow-policy`\n\n#### Description\nThe tree growing policy. Defines how to perform greedy tree construction.\n\nPossible values:\n\n- SymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric.\n- Depthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement.\n  Note\n  \n  Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\n- Lossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split.\n  Note\n  \n  Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\n\n**Type**\n\nstring\n\n**Default value**\n\nSymmetricTree\n\n**Supported processing units**\n\nCPU and GPU\n\n## min\\_data\\_in\\_leaf\nCommand-line: `--min-data-in-leaf`\n\n*Alias:* `min_child_samples`\n\n#### Description\nThe minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value.  \n Can be used only with the Lossguide and Depthwise growing policies.\n\n**Type**\n\nint\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU and GPU\n\n## max\\_leaves\nCommand-line: `--max-leaves`\n\n*Alias:*`num_leaves`\n\n#### Description\nThe maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy.\n\nNote\n\nIt is not recommended to use values greater than 64, since it can significantly slow down the training process.\n\n**Type**\n\nint\n\n**Default value**\n\n31\n\n**Supported processing units**\n\nCPU and GPU\n\n## ignored\\_features\nCommand-line: `-I`, `--ignore-features`\n\n#### Description\nFeature indices to exclude from the training.\n\nPython package\n\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices.\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored: `[1,2,7,42,43,44,45]`\n\nR package\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45).\n\nCommand-line\n\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices.\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction: `1:2:7:42-45`.\n\n**Default value**\n\nPython package, R package\n\nNone\n\nCommand-line\n\nOmitted\n\n**Supported processing units**\n\nCPU and GPU\n\n## one\\_hot\\_max\\_size\nCommand-line: `--one-hot-max-size`\n\n#### Description\nUse one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features.\n\nSee [details](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/features\/categorical-features).\n\n**Type**\n\nint\n\n**Default value**\n\nThe default value depends on various conditions:\n\n- N\/A if training is performed on CPU in Pairwise scoring mode\n  Read more about Pairwise scoring\n  \n  The following loss functions use Pairwise scoring:\n  \n  - YetiRankPairwise\n  - PairLogitPairwise\n  - QueryCrossEntropy\n  \n  Pairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions.\n- 255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training\n- 10 if training is performed in [Ranking](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/loss-functions-ranking) mode\n- 2 if none of the conditions above is met\n\n**Supported processing units**\n\nCPU and GPU\n\n## has\\_time\nCommand-line: `--has-time`\n\n#### Description\nUse the order of objects in the input data (do not perform random permutations during the [Transforming categorical features to numerical features](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-main-stages_cat-to-numberic) and [Choosing the tree structure](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-main-stages_choose-tree-structure) stages).\n\nThe Timestamp column type is used to determine the order of objects if specified in the [input data](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data).\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse (not used; generates random permutations)\n\n**Supported processing units**\n\nCPU and GPU\n\n## rsm\nCommand-line: `--rsm`\n\n*Alias:*`colsample_bylevel`\n\n#### Description\nRandom subspace method. The percentage of features to use at each split selection, when features are selected over again at random.\n\nThe value must be in the range (0;1\\].\n\n**Type**\n\nfloat (0;1\\]\n\n**Default value**\n\nNone (set to 1)\n\n**Supported processing units**\n\nCPU; GPU for pairwise ranking\n\n## nan\\_mode\nCommand-line: `--nan-mode`\n\n#### Description\nThe method for [processing missing values](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-missing-values-processing) in the input dataset.\n\nPossible values:\n\n- \"Forbidden\" — Missing values are not supported, their presence is interpreted as an error.\n- \"Min\" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\n- \"Max\" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\n\nUsing the Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree.\n\nNote\n\nThe method for processing missing values can be set individually for each feature in the [Custom quantization borders and missing value modes](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data_custom-borders) input file. Such values override the ones specified in this parameter.\n\n**Type**\n\nstring\n\n**Default value**\n\nMin\n\n**Supported processing units**\n\nCPU and GPU\n\n## input\\_borders\nCommand-line: `--input-borders-file`\n\n#### Description\nLoad [Custom quantization borders and missing value modes](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/input-data_custom-borders) from a file (do not generate them).\n\nBorders are automatically generated before training if this parameter is not set.\n\n**Type**\n\nstring\n\n**Default value**\n\nPython package\n\nNone\n\nCommand-line\n\nThe file is not loaded, the values are generated\n\n**Supported processing units**\n\nCPU and GPU\n\n## output\\_borders\nCommand-line: `--output-borders-file`\n\n#### Description\nSave quantization borders for the current dataset to a file.\n\nRefer to the [file format description](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/output-data_custom-borders).\n\n**Type**\n\nstring\n\n**Default value**\n\nPython package\n\nNone\n\nCommand-line\n\nThe file is not saved\n\n**Supported processing units**\n\nCPU and GPU\n\n## fold\\_permutation\\_block\nCommand-line: `--fold-permutation-block`\n\n#### Description\nObjects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\n1\n\nR package, Command-line\n\nDefault value differs depending on the dataset size and ranges from 1 to 256 inclusively\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_method\nCommand-line: `--leaf-estimation-method`\n\n#### Description\nThe method used to calculate the values in leaves.\n\nPossible values:\n\n- Newton\n- Gradient\n- Exact\n\n**Type**\n\nstring\n\n**Default value**\n\nDepends on the mode and the selected loss function:\n\n- Regression with Quantile or MAE loss functions — One Exact iteration.\n- Regression with any loss function but Quantile or MAE – One Gradient iteration.\n- Classification mode – Ten Newton iterations.\n- Multiclassification mode – One Newton iteration.\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_iterations\nCommand-line: `--leaf-estimation-iterations`\n\n#### Description\nCatBoost might calculate leaf values using several gradient or newton steps instead of a single one.\n\nThis parameter regulates how many steps are done in every tree when calculating leaf values.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\nNone (Depends on the training objective)\n\nR package, Command-line\n\nDepends on the training objective\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_backtracking\nCommand-line: `--leaf-estimation-backtracking`\n\n#### Description\nWhen the value of the `leaf_estimation_iterations` parameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree.\n\nThe behaviour differs depending on the value of this parameter:\n\n- No — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf.\n- Any other value —Backtracking is used.  \n   In this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf.\n\nWhen `leaf_estimation_iterations` for the Command-line version is set to `n`, the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having `n` gradient steps, the algorithm makes a single gradient step that is reduced `n` times, which means that it is divided by 2 ⋅ n 2\\\\cdot n 2⋅n times.\n\nPossible values:\n\n- No — Do not use backtracking. Supported on CPU and GPU.\n- AnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU.\n- Armijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU.\n\n**Type**\n\nstring\n\n**Default value**\n\nAnyImprovement\n\n**Supported processing units**\n\nDepends on the selected value\n\n## fold\\_len\\_multiplier\nCommand-line: `--fold-len-multiplier`\n\n#### Description\nCoefficient for changing the length of folds.\n\nThe value must be greater than 1. The best validation result is achieved with minimum values.\n\nWith values close to 1 (for example, 1 \\+ ϵ 1+\\\\epsilon 1\\+ϵ), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects.\n\n**Type**\n\nfloat\n\n**Default value**\n\n2\n\n**Supported processing units**\n\nCPU and GPU\n\n## approx\\_on\\_full\\_history\nCommand-line:`--approx-on-full-history`\n\n#### Description\nThe principles for calculating the approximated values.\n\nPossible values:\n\n- \"False\" — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows:\n  1\n  \n  X\n  \n  \\\\frac{1}X\n  \n  X1\n  , where `X` is the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate\n- \"True\" — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate.\n\n**Type**\n\nbool\n\n**Default value**\n\nPython package, Command-line\n\nFalse\n\nR package\n\nTrue\n\n**Supported processing units**\n\nCPU\n\n## class\\_weights\nCommand-line: `--class-weights`\n\n#### Description\nClass weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems.\n\nPython package\n\nNote\n\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right) (sum\\_positivesum\\_negative) for class 1.\n\nFor example, `class_weights=[0.1, 4]`multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\n\nIf class labels are not standard consecutive integers \\[0, 1 ... class\\_count-1\\], use the dict or collections.OrderedDict type with label to weight mapping.\n\nFor example, `class_weights={'a': 1.0, 'b': 0.5, 'c': 2.0}` multiplies the weights of objects with class label `a` by 1.0, the weights of objects with class label `b` by 0.5 and the weights of objects with class label `c` by 2.0.\n\nThe dictionary form can also be used with standard consecutive integers class labels for additional readability. For example: `class_weights={0: 1.0, 1: 0.5, 2: 2.0}`.\n\nNote\n\nClass labels are extracted from dictionary keys for the following types of class\\_weights:\n\n- dict\n- collections.OrderedDict (when the order of classes in the model is important)\n\nThe class\\_names parameter can be skipped when using these types.\n\nAlert\n\nDo not use this parameter with auto\\_class\\_weights and scale\\_pos\\_weight.\n\nR package\n\nFor example, `class_weights <- c(0.1, 4)` multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\n\nAlert\n\nDo not use this parameter with auto\\_class\\_weights.\n\nCommand-line\n\nNote\n\nThe quantity of class weights must match the quantity of class names specified in the `--class-names` parameter and the number of classes specified in the `--classes-count parameter`.\n\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right) (sum\\_positivesum\\_negative) for class 1.\n\nFormat:\n```\n<value for class 1>,..,<values for class N>\n```\nFor example:\n```\n0.85,1.2,1\n```\nAlert\n\nDo not use this parameter with auto\\_class\\_weights.\n\n**Type**\n\n- list\n- dict\n- collections.OrderedDict\n\n**Default value**\n\nNone (the weight for all classes is set to 1)\n\n**Supported processing units**\n\nCPU and GPU\n\n## class\\_names\n#### Description\nClasses names. Allows to redefine the default values when using the MultiClass and Logloss metrics.\n\nIf the upper limit for the numeric class label is specified, the number of classes names should match this value.\n\nWarning\n\nThe quantity of classes names must match the quantity of classes weights specified in the `--class-weights` parameter and the number of classes specified in the `--classes-count` parameter.\n\nFormat:\n```\n<name for class 1>,..,<name for class N>\n```\nFor example:\n```\nsmartphone,touchphone,tablet\n```\n**Type**\n\nlist of strings\n\n**Default value**\n\nNone\n\n**Supported processing units**\n\nCPU and GPU\n\n## auto\\_class\\_weights\nCommand-line: `--auto-class-weights`\n\n#### Description\nAutomatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights.\n\nSupported values:\n\n- None — All class weights are set to 1\n- Balanced:\n  C W k \\= m a x c \\= 1 K ( ∑ t i \\= c w i ) ∑ t i \\= k w i CW\\_k=\\\\displaystyle\\\\frac{max\\_{c=1}^K(\\\\sum\\_{t\\_{i}=c}{w\\_i})}{\\\\sum\\_{t\\_{i}=k}{w\\_{i}}} CWk\\=∑ti\\=kwimaxc\\=1K(∑ti\\=cwi)\n- SqrtBalanced:\n  C W k \\= m a x c \\= 1 K ( ∑ t i \\= c w i ) ∑ t i \\= k w i CW\\_k=\\\\sqrt{\\\\displaystyle\\\\frac{max\\_{c=1}^K(\\\\sum\\_{t\\_i=c}{w\\_i})}{\\\\sum\\_{t\\_i=k}{w\\_i}}} CWk\\= ∑ti\\=kwimaxc\\=1K(∑ti\\=cwi) \n\nAlert\n\nDo not use this parameter with `class_weights` and `scale_pos_weight`.\n\n**Type**\n\nstring\n\n**Default value**\n\nNone — All class weights are set to 1\n\n**Supported processing units**\n\nCPU and GPU\n\n## scale\\_pos\\_weight\n#### Description\nThe weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1.\n\nNote\n\nFor imbalanced datasets, the weight multiplier can be set to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right) (sum\\_positivesum\\_negative)\n\nAlert\n\nDo not use this parameter with `auto_class_weights` and `class_weights`.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\\.0\n\n**Supported processing units**\n\nCPU and GPU\n\n## boosting\\_type\nCommand-line: `--boosting-type`\n\n#### Description\nBoosting scheme.\n\nPossible values:\n\n- Ordered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme.\n- Plain — The classic gradient boosting scheme.\n\n**Type**\n\nstring\n\n**Default value**\n\nDepends on the processing unit type, the number of objects in the training dataset and the selected learning mode\n\n- CPU\n  Plain\n- GPU\n  - Any number of objects, MultiClass or MultiClassOneVsAll mode: Plain\n  - More than 50 thousand objects, any mode: Plain\n  - Less than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered\n\n**Supported processing units**\n\nCPU and GPU\n\nOnly the Plain mode is supported for the MultiClass loss on GPU\n\n## boost\\_from\\_average\nCommand-line: `--boost-from-average`\n\n#### Description\nInitialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value.\n\nAvailable for the following loss functions:\n\n- RMSE\n- Logloss\n- CrossEntropy\n- Quantile\n- MAE\n- MAPE\n\n**Type**\n\nbool\n\n**Default value**\n\nDepends on the selected loss function:\n\n- True for RMSE, Quantile, MAE, MAPE\n- False for all other loss functions\n\n**Supported processing units**\n\nCPU and GPU\n\n## langevin\nCommand-line: `--langevin`\n\n#### Description\nEnables the Stochastic Gradient Langevin Boosting mode.\n\nRefer to the [SGLB: Stochastic Gradient Langevin Boosting](https:\/\/arxiv.org\/abs\/2001.07248) paper for details.\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU\n\n## diffusion\\_temperature\nCommand-line: `--diffusion-temperature`\n\n#### Description\nThe diffusion temperature of the Stochastic Gradient Langevin Boosting mode.\n\nOnly non-negative values are supported.\n\n**Type**\n\nfloat\n\n**Default value**\n\n10000\n\n**Supported processing units**\n\nCPU\n\n## posterior\\_sampling\nCommand-line: `--posterior-sampling`\n\n#### Description\nIf this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties.  \n Specifies options:\n\n- `Langevin`: true,\n- `DiffusionTemperature`: objects in learn pool count,\n- `ModelShrinkRate`: 1 \/ (2. \\* objects in learn pool count).\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU only\n\n## allow\\_const\\_label\nCommand-line: `--allow-const-label`\n\n#### Description\nUse it to train models with datasets that have equal label values for all objects.\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU and GPU\n\n## score\\_function\nCommand-line: `--score-function`\n\n#### Description\nThe [score type](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-score-functions) used to select the next split during the tree construction.\n\nPossible values:\n\n- Cosine (do not use this score type with the Lossguide tree growing policy)\n- L2\n- NewtonCosine (do not use this score type with the Lossguide tree growing policy)\n- NewtonL2\n\n**Type**\n\nstring\n\n**Default value**\n\nCosine\n\n**Supported processing units**\n\nThe supported score functions vary depending on the processing unit type:\n\n- GPU — All score types\n- CPU — Cosine, L2\n## monotone\\_constraints\nCommand-line: `--monotone-constraints`\n\n#### Description\nImpose monotonic constraints on numerical features.\n\nPossible values:\n\n- \"1\" — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features.\n- \"\\-1\" — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features.\n- \"0\" — constraints are disabled.\n\nSupported formats for setting the value of this parameter (all feature indices are zero-based):\n\n- Set constraints individually for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<constraint_0>, <constraint_2>, .., <constraint_n-1>)\"\n  ```\n  Zero constraints for features at the end of the list may be dropped.\n  In `monotone_constraints = \"(1,0,-1)\"`an increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features.\n- Set constraints individually for each explicitly specified feature as a string (the number of features is n).\n  ```\n  \"<feature index or name>:<constraint>, .., <feature index or name>:<constraint>\"\n  ```\n  These examples\n  ```\n  monotone-constraints = \"2:1,4:-1\"\n  ```\n  ```\n  monotone-constraints = \"Feature2:1,Feature4:-1\"\n  ```\n  are identical, given that the name of the feature index 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set constraints individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<constraint_0>, <constraint_2>, .., <constraint_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<constraint>, .., \"<feature index or name>\":<constraint>}\n  ```\n  Array examples\n  ```\n  monotone_constraints = [1, 0, -1]\n  ```\n  These dictionary examples\n  \n  ```\n  monotone_constraints = {\"Feature2\":1,\"Feature4\":-1}\n  ```\n  ```\n  monotone_constraints = {\"2\":1, \"4\":-1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list of strings\n- string\n- dict\n- list\n\n**Default value**\n\nPython package, R package\n\nNone\n\nCommand-line\n\nOmmited\n\n**Supported processing units**\n\nCPU\n\n## feature\\_weights\nCommand-line: `--feature-weights`\n\n#### Description\nPer-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split.\n\nNon-negative float values are supported for each weight.\n\nSupported formats for setting the value of this parameter:\n\n- Set the multiplication weight for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped.\n  In this\n  example\n  ```\n  feature_weights = \"(0.1,1,3)\"\n  ```\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the multiplication weight individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<weight>, .., <feature index or name>:<weight>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples\n  \n  ```\n  feature_weights = \"2:1.1,4:0.1\"\n  ```\n  ```\n  feature_weights = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<weight>, .., \"<feature index or name>\":<weight>}\n  ```\n  Array examples\n  ```\n  feature_weights = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  \n  ```\n  feature_weights = {\"Feature2\":1.1,\"Feature4\":0.3}\n  ```\n  ```\n  feature_weights = {\"2\":1.1, \"4\":0.3}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n1 for all features\n\n**Supported processing units**\n\nCPU\n\n## first\\_feature\\_use\\_penalties\nCommand-line: `--first-feature-use-penalties`\n\n#### Description\nPer-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model.\n\nRefer to the [Per-object and per-feature penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-score-functions) section for details on applying different score penalties.\n\nNon-negative float values are supported for each penalty.\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  In this example\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the penalty individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\n  ```\n  Array examples.\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = [0.1, 1, 3]\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  first_feature_use_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  per_object_feature_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n0 for all features\n\n**Supported processing units**\n\nCPU\n\n## fixed\\_binary\\_splits\nCommand-line: `--fixed-binary-splits`\n\n#### Description\nA list of indices of binary features to put at the top of each tree; ignored if `grow_policy` is `Symmetric`.\n\n**Type**\n\nlist\n\n**Default value**\n\nNone\n\n**Supported processing units**\n\nGPU\n\n## penalties\\_coefficient\nCommand-line: `--penalties-coefficient`\n\n#### Description\nA single-value common coefficient to multiply all penalties.\n\nNon-negative values are supported.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU\n\n## per\\_object\\_feature\\_penalties\nCommand-line: `--per-object-feature-penalties`\n\n#### Description\nPer-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time.\n\nRefer to the [Per-object and per-feature penalties](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/concepts\/algorithm-score-functions) section for details on applying different score penalties.\n\nNon-negative float values are supported for each penalty.\n\nPython package\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  In this example\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the penalty individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\n  ```\n  Array examples.\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = [0.1, 1, 3]\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  first_feature_use_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  per_object_feature_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\nR package\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  Penalties equal to 0 at the end of the list may be dropped.\n  In this\n  example\n  \n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples\n  \n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n0 for all objects\n\n**Supported processing units**\n\nCPU\n\n## model\\_shrink\\_rate\nCommand-line: `--model-shrink-rate`\n\n#### Description\nThe constant used to calculate the coefficient for multiplying the model on each iteration.  \n The actual model shrinkage coefficient calculated at each iteration depends on the value of the `--model-shrink-mode`for the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1\\].\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value depends on the values of the following parameters:\n\n- `--model-shrink-mode` for the Command-line version\n- `--monotone-constraints` for the Command-line version\n\n**Supported processing units**\n\nCPU\n\n## model\\_shrink\\_mode\nCommand-line: `model_shrink_mode`\n\n#### Description\nDetermines how the actual model shrinkage coefficient is calculated at each iteration.\n\nPossible values:\n\n- Constant:\n  1 − m o d e l \\_ s h r i n k \\_ r a t e ⋅ l e a r n i n g \\_ r a t e , 1 - model\\\\\\_shrink\\\\\\_rate \\\\cdot learning\\\\\\_rate {,} 1−model\\_shrink\\_rate⋅learning\\_rate,\n  - m\n    \n    o\n    \n    d\n    \n    e\n    \n    l\n    \n    \\_\n    \n    s\n    \n    h\n    \n    r\n    \n    i\n    \n    n\n    \n    k\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    model\\\\\\_shrink\\\\\\_rate\n    \n    model\\_shrink\\_rate\n    is the value of the `--model-shrink-rate`for the Command-line version parameter.\n  - l\n    \n    e\n    \n    a\n    \n    r\n    \n    n\n    \n    i\n    \n    n\n    \n    g\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    learning\\\\\\_rate\n    \n    learning\\_rate\n    is the value of the `--learning-rate`for the Command-line version parameter\n- Decreasing:\n  1 − m o d e l \\_ s h r i n k \\_ r a t e i , 1 - \\\\frac{model\\\\\\_shrink\\\\\\_rate}{i} {,} 1−imodel\\_shrink\\_rate,\n  - m\n    \n    o\n    \n    d\n    \n    e\n    \n    l\n    \n    \\_\n    \n    s\n    \n    h\n    \n    r\n    \n    i\n    \n    n\n    \n    k\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    model\\\\\\_shrink\\\\\\_rate\n    \n    model\\_shrink\\_rate\n    is the value of the `--model-shrink-rate`for the Command-line version parameter.\n  - i\n    \n    i\n    \n    i\n    is the identifier of the iteration.\n\n**Type**\n\nstring\n\n**Default value**\n\nConstant\n\n**Supported processing units**\n\nCPU\n\n### Was the article helpful?\nYes\n\nNo\n\nPrevious\n\n[Overview](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/)\n\nNext\n\n[CTR settings](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/en\/references\/training-parameters\/ctr)\n\n![](https:\/\/mc.yandex.ru\/watch\/60763294)","attrs_readable_markdown":"## loss\\_function\nCommand-line: `--loss-function`\n\n*Alias:* `objective`\n\n#### Description\nThe [metric](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions) to use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\nSupported metrics\n\n- RMSE\n- Logloss\n- MAE\n- CrossEntropy\n- Quantile\n- LogLinQuantile\n- Lq\n- MultiRMSE\n- MultiClass\n- MultiClassOneVsAll\n- MultiLogloss\n- MultiCrossEntropy\n- MAPE\n- Poisson\n- PairLogit\n- PairLogitPairwise\n- QueryRMSE\n- QuerySoftMax\n- GroupQuantile\n- Tweedie\n- YetiRank\n- YetiRankPairwise\n- StochasticFilter\n- StochasticRank\n\nA custom python object can also be set as the value of this parameter (see an [example](https:\/\/catboost.ai\/docs\/en\/concepts\/python-usages-examples)).\n\nFor example, use the following construction to calculate the value of Quantile with the coefficient α \\= 0\\.1 \\\\alpha = 0.1:\n```\nQuantile:alpha=0.1\n```\n**Type**\n\n- string\n- object\n\n**Default value**\n\nPython package\n\nDepends on the class:\n\n- [CatBoostClassifier](https:\/\/catboost.ai\/docs\/en\/concepts\/python-reference_catboostclassifier): Logloss if the `target_border` parameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass.\n- [CatBoost](https:\/\/catboost.ai\/docs\/en\/concepts\/python-reference_catboost) and [CatBoostRegressor](https:\/\/catboost.ai\/docs\/en\/concepts\/python-reference_catboostregressor): RMSE\n\nR package, Command-line\n\nRMSE\n\n**Supported processing units**\n\nCPU and GPU\n\n## custom\\_metric\nCommand-line: `--custom-metric`\n\n#### Description\n[Metric](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions) values to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\n[Supported metrics](https:\/\/catboost.ai\/docs\/en\/references\/custom-metric__supported-metrics)\n\nExamples\n\n- Calculate the value of CrossEntropy:\n  ```\n  CrossEntropy\n  ```\n- Calculate the value of Quantile with the coefficient α \\= 0\\.1 \\\\alpha = 0.1\n  ```\n  Quantile:alpha=0.1\n  ```\n- Calculate the values of Logloss and AUC:\n  ```\n  ['Logloss', 'AUC']\n  ```\n\nValues of all custom metrics for learn and validation datasets are saved to the [Metric](https:\/\/catboost.ai\/docs\/en\/concepts\/output-data_loss-function) output files (`learn_error.tsv` and `test_error.tsv` respectively). The directory for these files is specified in the `--train-dir` (`train_dir`) parameter.\n\nUse the [visualization tools](https:\/\/catboost.ai\/docs\/en\/features\/visualization) to see a live chart with the dynamics of the specified metrics.\n\n**Type**\n\n- string\n- list of strings\n\n**Default value**\n\nPython package\n\nNone\n\nR package\n\nNone\n\nCommand-line\n\nNone (do not output additional metric values)\n\n**Supported processing units**\n\nCPU and GPU\n\n## eval\\_metric\nCommand-line: `--eval-metric`\n\n#### Description\nThe metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the [Objectives and metrics](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions) section for details on each metric).\n\nFormat:\n```\n<Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>]\n```\n[Supported metrics](https:\/\/catboost.ai\/docs\/en\/references\/eval-metric__supported-metrics)\n\nA user-defined function can also be set as the value (see an [example](https:\/\/catboost.ai\/docs\/en\/concepts\/python-usages-examples)).\n\nExamples:\n```\nR2\n```\n**Type**\n\n- string\n- object\n\n**Default value**\n\nOptimized objective is used\n\n**Supported processing units**\n\nCPU and GPU\n\n## iterations\nCommand-line: `-i`, `--iterations`\n\n*Aliases:* `num_boost_round`, `n_estimators`, `num_trees`\n\n#### Description\nThe maximum number of trees that can be built when solving machine learning problems.\n\nWhen using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.\n\n**Type**\n\nint\n\n**Default value**\n\n1000\n\n**Supported processing units**\n\nCPU and GPU\n\n## learning\\_rate\nCommand-line: `-w`, `--learning-rate`\n\n*Alias:* `eta`\n\n#### Description\nThe learning rate.\n\nUsed for reducing the gradient step.\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value is defined automatically for [`Logloss`](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions-classification#Logit), [`MultiClass`](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions-multiclassification#MultiClass) and [`RMSE`](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions-regression#RMSE) loss functions depending on the number of iterations if none of parameters [`leaf_estimation_iterations`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/common#leaf_estimation_iterations), [`leaf_estimation_method`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/common#leaf_estimation_method), [`l2_leaf_reg`](https:\/\/catboost.ai\/docs\/en\/references\/training-parameters\/common#l2_leaf_reg) is set. In this case, the selected learning rate is printed to stdout and saved in the model.\n\nIn other cases, the default value is 0.03.\n\n**Supported processing units**\n\nCPU and GPU\n\n## random\\_seed\nCommand-line: `-r`, `--random-seed`\n\n*Alias:*`random_state`\n\n#### Description\nThe random seed used for training.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\nNone (0)\n\nR package, Command-line\n\n0\n\n**Supported processing units**\n\nCPU and GPU\n\n## l2\\_leaf\\_reg\nCommand-line: `--l2-leaf-reg`, `l2-leaf-regularizer`\n\n*Alias:* `reg_lambda`\n\n#### Description\nCoefficient at the L2 regularization term of the cost function.\n\nAny positive value is allowed.\n\n**Type**\n\nfloat\n\n**Default value**\n\n3\\.0\n\n**Supported processing units**\n\nCPU and GPU\n\n## bootstrap\\_type\nCommand-line: `--bootstrap-type`\n\n#### Description\n[Bootstrap type](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-main-stages_bootstrap-options). Defines the method for sampling the weights of objects.\n\nSupported methods:\n\n- Bayesian\n- Bernoulli\n- MVS\n- Poisson (supported for GPU only)\n- No\n\n**Type**\n\nstring\n\n**Default value**\n\nThe default value depends on `objective`, `task_type`, `bagging_temperature` and `sampling_unit`:\n\n- When the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging\\_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5.\n- Neither MultiClass nor MultiClassOneVsAll, task\\_type = CPU and sampling\\_unit = Object: MVS with the subsample parameter set to 0.8.\n- Otherwise: Bayesian.\n\n**Supported processing units**\n\nCPU and GPU\n\n## bagging\\_temperature\nCommand-line: `--bagging-temperature`\n\n#### Description\nDefines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes.\n\nUse the Bayesian bootstrap to assign random weights to objects.\n\nThe weights are sampled from exponential distribution if the value of this parameter is set to \"1\". All weights are equal to 1 if the value of this parameter is set to \"0\".\n\nPossible values are in the range \\[ 0 ; inf ⁡ ) \\[0; \\\\inf). The higher the value the more aggressive the bagging is.\n\nThis parameter can be used if the selected bootstrap type is Bayesian.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU and GPU\n\n## subsample\nCommand-line: `--subsample`\n\n#### Description\nSample rate for bagging.\n\nThis parameter can be used if one of the following bootstrap types is selected:\n\n- Poisson\n- Bernoulli\n- MVS\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value depends on the dataset size and the bootstrap type:\n\n- Datasets with less than 100 objects — 1\n- Datasets with 100 objects or more:\n  - Poisson, Bernoulli — 0.66\n  - MVS — 0.8\n\n**Supported processing units**\n\nCPU and GPU\n\n## sampling\\_frequency\nCommand-line: `--sampling-frequency`\n\n#### Description\nFrequency to sample weights and objects when building trees.\n\nSupported values:\n\n- PerTree — Before constructing each new tree\n- PerTreeLevel — Before choosing each new split of a tree\n\n**Type**\n\nstring\n\n**Default value**\n\nPerTreeLevel\n\n**Supported processing units**\n\nCPU\n\n## sampling\\_unit\nCommand-line: `--sampling-unit`\n\n#### Description\nThe sampling scheme.\n\nPossible values:\n\n- Object — The weight\n  w\n  \n  i\n  \n  w\\_{i}\n  of the i-th object\n  o\n  \n  i\n  \n  o\\_{i}\n  is used for sampling the corresponding object.\n- Group — The weight\n  w\n  \n  j\n  \n  w\\_{j}\n  of the group\n  g\n  \n  j\n  \n  g\\_{j}\n  is used for sampling each object\n  o\n  \n  i\n  \n  j\n  \n  o\\_{i\\_{j}}\n  from the group\n  g\n  \n  j\n  \n  g\\_{j}\n  .\n\n**Type**\n\nString\n\n**Default value**\n\nObject\n\n**Supported processing units**\n\nCPU and GPU\n\n## mvs\\_reg\nCommand-line: `--mvs-reg`\n\n#### Description\nAffects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to ∞ \\\\infty - Bernoulli).\n\nNote\n\nThis parameter is supported only for the MVS sampling method (the `bootstrap_type` parameter must be set to MVS).\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe value is set based on the gradient distribution on the current iteration\n\n**Supported processing units**\n\nCPU\n\n## random\\_strength\nCommand-line: `--random-strength`\n\n#### Description\nThe amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model.\n\nThe value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected.\n\nThe scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance.\n\nNote\n\nThis parameter is not supported for the following loss functions:\n\n- QueryCrossEntropy\n- YetiRankPairwise\n- PairLogitPairwise\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU\n\n## use\\_best\\_model\nCommand-line: `--use-best-model`\n\n#### Description\nIf this parameter is set, the number of trees that are saved in the resulting model is defined as follows:\n\n1. Build the number of trees defined by the training parameters.\n2. Use the validation dataset to identify the iteration with the optimal value of the metric specified in `--eval-metric` (`--eval-metric`).\n\nNo trees are saved after this iteration.\n\nThis option requires a validation dataset to be provided.\n\n**Type**\n\nbool\n\n**Default value**\n\nTrue if a validation set is input (the eval\\_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise.\n\n**Supported processing units**\n\nCPU and GPU\n\n## best\\_model\\_min\\_trees\nCommand-line: `--best-model-min-trees`\n\n#### Description\nThe minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees.\n\nShould be used with the `--use-best-model` parameter.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package, R package\n\nNone (The minimal number of trees for the best model is not set)\n\nCommand-line\n\nThe minimal number of trees for the best model is not set\n\n**Supported processing units**\n\nCPU and GPU\n\n## depth\nCommand-line: `-n`, `--depth`\n\n*Alias:* `max_depth`\n\n#### Description\nDepth of the trees.\n\nThe range of supported values depends on the processing unit type and the type of the selected loss function:\n\n- CPU — Any integer up to 16.\n- GPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions.\n\n**Type**\n\nint\n\n**Default value**\n\n6 (16 if the growing policy is set to Lossguide)\n\n**Supported processing units**\n\nCPU and GPU\n\n## grow\\_policy\nCommand-line: `--grow-policy`\n\n#### Description\nThe tree growing policy. Defines how to perform greedy tree construction.\n\nPossible values:\n\n- SymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric.\n- Depthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement.\n  Note\n  \n  Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\n- Lossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split.\n  Note\n  \n  Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm.\n\n**Type**\n\nstring\n\n**Default value**\n\nSymmetricTree\n\n**Supported processing units**\n\nCPU and GPU\n\n## min\\_data\\_in\\_leaf\nCommand-line: `--min-data-in-leaf`\n\n*Alias:* `min_child_samples`\n\n#### Description\nThe minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value.  \n Can be used only with the Lossguide and Depthwise growing policies.\n\n**Type**\n\nint\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU and GPU\n\n## max\\_leaves\nCommand-line: `--max-leaves`\n\n*Alias:*`num_leaves`\n\n#### Description\nThe maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy.\n\nNote\n\nIt is not recommended to use values greater than 64, since it can significantly slow down the training process.\n\n**Type**\n\nint\n\n**Default value**\n\n31\n\n**Supported processing units**\n\nCPU and GPU\n\n## ignored\\_features\nCommand-line: `-I`, `--ignore-features`\n\n#### Description\nFeature indices to exclude from the training.\n\nPython package\n\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices.\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored: `[1,2,7,42,43,44,45]`\n\nR package\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45).\n\nCommand-line\n\nIt is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices.\n\nSpecifics:\n\n- Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to \"42\", the corresponding non-existing feature is successfully ignored.\n- The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https:\/\/catboost.ai\/docs\/en\/concepts\/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\\t>label value<\\t>num feature`. So for the row `rock<\\t>0<\\t>42`, the identifier for the \"rock\" feature is 0, and for the \"42\" feature it's 1.\n\nFor example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction: `1:2:7:42-45`.\n\n**Default value**\n\nPython package, R package\n\nNone\n\nCommand-line\n\nOmitted\n\n**Supported processing units**\n\nCPU and GPU\n\n## one\\_hot\\_max\\_size\nCommand-line: `--one-hot-max-size`\n\n#### Description\nUse one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features.\n\nSee [details](https:\/\/catboost.ai\/docs\/en\/features\/categorical-features).\n\n**Type**\n\nint\n\n**Default value**\n\nThe default value depends on various conditions:\n\n- N\/A if training is performed on CPU in Pairwise scoring mode\n  Read more about Pairwise scoring\n  \n  The following loss functions use Pairwise scoring:\n  \n  - YetiRankPairwise\n  - PairLogitPairwise\n  - QueryCrossEntropy\n  \n  Pairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions.\n- 255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training\n- 10 if training is performed in [Ranking](https:\/\/catboost.ai\/docs\/en\/concepts\/loss-functions-ranking) mode\n- 2 if none of the conditions above is met\n\n**Supported processing units**\n\nCPU and GPU\n\n## has\\_time\nCommand-line: `--has-time`\n\n#### Description\nUse the order of objects in the input data (do not perform random permutations during the [Transforming categorical features to numerical features](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-main-stages_cat-to-numberic) and [Choosing the tree structure](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-main-stages_choose-tree-structure) stages).\n\nThe Timestamp column type is used to determine the order of objects if specified in the [input data](https:\/\/catboost.ai\/docs\/en\/concepts\/input-data).\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse (not used; generates random permutations)\n\n**Supported processing units**\n\nCPU and GPU\n\n## rsm\nCommand-line: `--rsm`\n\n*Alias:*`colsample_bylevel`\n\n#### Description\nRandom subspace method. The percentage of features to use at each split selection, when features are selected over again at random.\n\nThe value must be in the range (0;1\\].\n\n**Type**\n\nfloat (0;1\\]\n\n**Default value**\n\nNone (set to 1)\n\n**Supported processing units**\n\nCPU; GPU for pairwise ranking\n\n## nan\\_mode\nCommand-line: `--nan-mode`\n\n#### Description\nThe method for [processing missing values](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-missing-values-processing) in the input dataset.\n\nPossible values:\n\n- \"Forbidden\" — Missing values are not supported, their presence is interpreted as an error.\n- \"Min\" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\n- \"Max\" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees.\n\nUsing the Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree.\n\n**Type**\n\nstring\n\n**Default value**\n\nMin\n\n**Supported processing units**\n\nCPU and GPU\n\n## input\\_borders\nCommand-line: `--input-borders-file`\n\n#### Description\nLoad [Custom quantization borders and missing value modes](https:\/\/catboost.ai\/docs\/en\/concepts\/input-data_custom-borders) from a file (do not generate them).\n\nBorders are automatically generated before training if this parameter is not set.\n\n**Type**\n\nstring\n\n**Default value**\n\nPython package\n\nNone\n\nCommand-line\n\nThe file is not loaded, the values are generated\n\n**Supported processing units**\n\nCPU and GPU\n\n## output\\_borders\nCommand-line: `--output-borders-file`\n\n#### Description\nSave quantization borders for the current dataset to a file.\n\nRefer to the [file format description](https:\/\/catboost.ai\/docs\/en\/concepts\/output-data_custom-borders).\n\n**Type**\n\nstring\n\n**Default value**\n\nPython package\n\nNone\n\nCommand-line\n\nThe file is not saved\n\n**Supported processing units**\n\nCPU and GPU\n\n## fold\\_permutation\\_block\nCommand-line: `--fold-permutation-block`\n\n#### Description\nObjects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\n1\n\nR package, Command-line\n\nDefault value differs depending on the dataset size and ranges from 1 to 256 inclusively\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_method\nCommand-line: `--leaf-estimation-method`\n\n#### Description\nThe method used to calculate the values in leaves.\n\nPossible values:\n\n- Newton\n- Gradient\n- Exact\n\n**Type**\n\nstring\n\n**Default value**\n\nDepends on the mode and the selected loss function:\n\n- Regression with Quantile or MAE loss functions — One Exact iteration.\n- Regression with any loss function but Quantile or MAE – One Gradient iteration.\n- Classification mode – Ten Newton iterations.\n- Multiclassification mode – One Newton iteration.\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_iterations\nCommand-line: `--leaf-estimation-iterations`\n\n#### Description\nCatBoost might calculate leaf values using several gradient or newton steps instead of a single one.\n\nThis parameter regulates how many steps are done in every tree when calculating leaf values.\n\n**Type**\n\nint\n\n**Default value**\n\nPython package\n\nNone (Depends on the training objective)\n\nR package, Command-line\n\nDepends on the training objective\n\n**Supported processing units**\n\nCPU and GPU\n\n## leaf\\_estimation\\_backtracking\nCommand-line: `--leaf-estimation-backtracking`\n\n#### Description\nWhen the value of the `leaf_estimation_iterations` parameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree.\n\nThe behaviour differs depending on the value of this parameter:\n\n- No — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf.\n- Any other value —Backtracking is used.  \n   In this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf.\n\nWhen `leaf_estimation_iterations` for the Command-line version is set to `n`, the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having `n` gradient steps, the algorithm makes a single gradient step that is reduced `n` times, which means that it is divided by 2 ⋅ n 2\\\\cdot n times.\n\nPossible values:\n\n- No — Do not use backtracking. Supported on CPU and GPU.\n- AnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU.\n- Armijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU.\n\n**Type**\n\nstring\n\n**Default value**\n\nAnyImprovement\n\n**Supported processing units**\n\nDepends on the selected value\n\n## fold\\_len\\_multiplier\nCommand-line: `--fold-len-multiplier`\n\n#### Description\nCoefficient for changing the length of folds.\n\nThe value must be greater than 1. The best validation result is achieved with minimum values.\n\nWith values close to 1 (for example, 1 \\+ ϵ 1+\\\\epsilon), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects.\n\n**Type**\n\nfloat\n\n**Default value**\n\n2\n\n**Supported processing units**\n\nCPU and GPU\n\n## approx\\_on\\_full\\_history\nCommand-line:`--approx-on-full-history`\n\n#### Description\nThe principles for calculating the approximated values.\n\nPossible values:\n\n- \"False\" — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows:\n  1\n  \n  X\n  \n  \\\\frac{1}X\n  , where `X` is the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate\n- \"True\" — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate.\n\n**Type**\n\nbool\n\n**Default value**\n\nPython package, Command-line\n\nFalse\n\nR package\n\nTrue\n\n**Supported processing units**\n\nCPU\n\n## class\\_weights\nCommand-line: `--class-weights`\n\n#### Description\nClass weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems.\n\nPython package\n\nNote\n\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right) for class 1.\n\nFor example, `class_weights=[0.1, 4]`multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\n\nIf class labels are not standard consecutive integers \\[0, 1 ... class\\_count-1\\], use the dict or collections.OrderedDict type with label to weight mapping.\n\nFor example, `class_weights={'a': 1.0, 'b': 0.5, 'c': 2.0}` multiplies the weights of objects with class label `a` by 1.0, the weights of objects with class label `b` by 0.5 and the weights of objects with class label `c` by 2.0.\n\nThe dictionary form can also be used with standard consecutive integers class labels for additional readability. For example: `class_weights={0: 1.0, 1: 0.5, 2: 2.0}`.\n\nNote\n\nClass labels are extracted from dictionary keys for the following types of class\\_weights:\n\n- dict\n- collections.OrderedDict (when the order of classes in the model is important)\n\nThe class\\_names parameter can be skipped when using these types.\n\nAlert\n\nDo not use this parameter with auto\\_class\\_weights and scale\\_pos\\_weight.\n\nR package\n\nFor example, `class_weights <- c(0.1, 4)` multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4.\n\nAlert\n\nDo not use this parameter with auto\\_class\\_weights.\n\nCommand-line\n\nNote\n\nThe quantity of class weights must match the quantity of class names specified in the `--class-names` parameter and the number of classes specified in the `--classes-count parameter`.\n\nFor imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right) for class 1.\n\nFormat:\n```\n<value for class 1>,..,<values for class N>\n```\nFor example:\n```\n0.85,1.2,1\n```\nAlert\n\nDo not use this parameter with auto\\_class\\_weights.\n\n**Type**\n\n- list\n- dict\n- collections.OrderedDict\n\n**Default value**\n\nNone (the weight for all classes is set to 1)\n\n**Supported processing units**\n\nCPU and GPU\n\n## class\\_names\n#### Description\nClasses names. Allows to redefine the default values when using the MultiClass and Logloss metrics.\n\nIf the upper limit for the numeric class label is specified, the number of classes names should match this value.\n\nWarning\n\nThe quantity of classes names must match the quantity of classes weights specified in the `--class-weights` parameter and the number of classes specified in the `--classes-count` parameter.\n\nFormat:\n```\n<name for class 1>,..,<name for class N>\n```\nFor example:\n```\nsmartphone,touchphone,tablet\n```\n**Type**\n\nlist of strings\n\n**Default value**\n\nNone\n\n**Supported processing units**\n\nCPU and GPU\n\n## auto\\_class\\_weights\nCommand-line: `--auto-class-weights`\n\n#### Description\nAutomatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights.\n\nSupported values:\n\n- None — All class weights are set to 1\n- Balanced:\n  C W k \\= m a x c \\= 1 K ( ∑ t i \\= c w i ) ∑ t i \\= k w i CW\\_k=\\\\displaystyle\\\\frac{max\\_{c=1}^K(\\\\sum\\_{t\\_{i}=c}{w\\_i})}{\\\\sum\\_{t\\_{i}=k}{w\\_{i}}}\n- SqrtBalanced:\n  C W k \\= m a x c \\= 1 K ( ∑ t i \\= c w i ) ∑ t i \\= k w i CW\\_k=\\\\sqrt{\\\\displaystyle\\\\frac{max\\_{c=1}^K(\\\\sum\\_{t\\_i=c}{w\\_i})}{\\\\sum\\_{t\\_i=k}{w\\_i}}}\n\nAlert\n\nDo not use this parameter with `class_weights` and `scale_pos_weight`.\n\n**Type**\n\nstring\n\n**Default value**\n\nNone — All class weights are set to 1\n\n**Supported processing units**\n\nCPU and GPU\n\n## scale\\_pos\\_weight\n#### Description\nThe weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1.\n\nNote\n\nFor imbalanced datasets, the weight multiplier can be set to ( s u m \\_ n e g a t i v e s u m \\_ p o s i t i v e ) \\\\left(\\\\frac{sum\\\\\\_negative}{sum\\\\\\_positive}\\\\right)\n\nAlert\n\nDo not use this parameter with `auto_class_weights` and `class_weights`.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\\.0\n\n**Supported processing units**\n\nCPU and GPU\n\n## boosting\\_type\nCommand-line: `--boosting-type`\n\n#### Description\nBoosting scheme.\n\nPossible values:\n\n- Ordered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme.\n- Plain — The classic gradient boosting scheme.\n\n**Type**\n\nstring\n\n**Default value**\n\nDepends on the processing unit type, the number of objects in the training dataset and the selected learning mode\n\n- CPU\n  Plain\n- GPU\n  - Any number of objects, MultiClass or MultiClassOneVsAll mode: Plain\n  - More than 50 thousand objects, any mode: Plain\n  - Less than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered\n\n**Supported processing units**\n\nCPU and GPU\n\nOnly the Plain mode is supported for the MultiClass loss on GPU\n\n## boost\\_from\\_average\nCommand-line: `--boost-from-average`\n\n#### Description\nInitialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value.\n\nAvailable for the following loss functions:\n\n- RMSE\n- Logloss\n- CrossEntropy\n- Quantile\n- MAE\n- MAPE\n\n**Type**\n\nbool\n\n**Default value**\n\nDepends on the selected loss function:\n\n- True for RMSE, Quantile, MAE, MAPE\n- False for all other loss functions\n\n**Supported processing units**\n\nCPU and GPU\n\n## langevin\nCommand-line: `--langevin`\n\n#### Description\nEnables the Stochastic Gradient Langevin Boosting mode.\n\nRefer to the [SGLB: Stochastic Gradient Langevin Boosting](https:\/\/arxiv.org\/abs\/2001.07248) paper for details.\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU\n\n## diffusion\\_temperature\nCommand-line: `--diffusion-temperature`\n\n#### Description\nThe diffusion temperature of the Stochastic Gradient Langevin Boosting mode.\n\nOnly non-negative values are supported.\n\n**Type**\n\nfloat\n\n**Default value**\n\n10000\n\n**Supported processing units**\n\nCPU\n\n## posterior\\_sampling\nCommand-line: `--posterior-sampling`\n\n#### Description\nIf this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties.  \n Specifies options:\n\n- `Langevin`: true,\n- `DiffusionTemperature`: objects in learn pool count,\n- `ModelShrinkRate`: 1 \/ (2. \\* objects in learn pool count).\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU only\n\n## allow\\_const\\_label\nCommand-line: `--allow-const-label`\n\n#### Description\nUse it to train models with datasets that have equal label values for all objects.\n\n**Type**\n\nbool\n\n**Default value**\n\nFalse\n\n**Supported processing units**\n\nCPU and GPU\n\n## score\\_function\nCommand-line: `--score-function`\n\n#### Description\nThe [score type](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-score-functions) used to select the next split during the tree construction.\n\nPossible values:\n\n- Cosine (do not use this score type with the Lossguide tree growing policy)\n- L2\n- NewtonCosine (do not use this score type with the Lossguide tree growing policy)\n- NewtonL2\n\n**Type**\n\nstring\n\n**Default value**\n\nCosine\n\n**Supported processing units**\n\nThe supported score functions vary depending on the processing unit type:\n\n- GPU — All score types\n- CPU — Cosine, L2\n## monotone\\_constraints\nCommand-line: `--monotone-constraints`\n\n#### Description\nImpose monotonic constraints on numerical features.\n\nPossible values:\n\n- \"1\" — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features.\n- \"\\-1\" — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features.\n- \"0\" — constraints are disabled.\n\nSupported formats for setting the value of this parameter (all feature indices are zero-based):\n\n- Set constraints individually for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<constraint_0>, <constraint_2>, .., <constraint_n-1>)\"\n  ```\n  Zero constraints for features at the end of the list may be dropped.\n  In `monotone_constraints = \"(1,0,-1)\"`an increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features.\n- Set constraints individually for each explicitly specified feature as a string (the number of features is n).\n  ```\n  \"<feature index or name>:<constraint>, .., <feature index or name>:<constraint>\"\n  ```\n  These examples\n  ```\n  monotone-constraints = \"2:1,4:-1\"\n  ```\n  ```\n  monotone-constraints = \"Feature2:1,Feature4:-1\"\n  ```\n  are identical, given that the name of the feature index 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set constraints individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<constraint_0>, <constraint_2>, .., <constraint_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<constraint>, .., \"<feature index or name>\":<constraint>}\n  ```\n  Array examples\n  ```\n  monotone_constraints = [1, 0, -1]\n  ```\n  These dictionary examples\n  \n  ```\n  monotone_constraints = {\"Feature2\":1,\"Feature4\":-1}\n  ```\n  ```\n  monotone_constraints = {\"2\":1, \"4\":-1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list of strings\n- string\n- dict\n- list\n\n**Default value**\n\nPython package, R package\n\nNone\n\nCommand-line\n\nOmmited\n\n**Supported processing units**\n\nCPU\n\n## feature\\_weights\nCommand-line: `--feature-weights`\n\n#### Description\nPer-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split.\n\nNon-negative float values are supported for each weight.\n\nSupported formats for setting the value of this parameter:\n\n- Set the multiplication weight for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped.\n  In this\n  example\n  ```\n  feature_weights = \"(0.1,1,3)\"\n  ```\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the multiplication weight individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<weight>, .., <feature index or name>:<weight>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples\n  \n  ```\n  feature_weights = \"2:1.1,4:0.1\"\n  ```\n  ```\n  feature_weights = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<weight>, .., \"<feature index or name>\":<weight>}\n  ```\n  Array examples\n  ```\n  feature_weights = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  \n  ```\n  feature_weights = {\"Feature2\":1.1,\"Feature4\":0.3}\n  ```\n  ```\n  feature_weights = {\"2\":1.1, \"4\":0.3}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n1 for all features\n\n**Supported processing units**\n\nCPU\n\n## first\\_feature\\_use\\_penalties\nCommand-line: `--first-feature-use-penalties`\n\n#### Description\nPer-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model.\n\nRefer to the [Per-object and per-feature penalties](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-score-functions) section for details on applying different score penalties.\n\nNon-negative float values are supported for each penalty.\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  In this example\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the penalty individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\n  ```\n  Array examples.\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = [0.1, 1, 3]\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  first_feature_use_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  per_object_feature_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n0 for all features\n\n**Supported processing units**\n\nCPU\n\n## fixed\\_binary\\_splits\nCommand-line: `--fixed-binary-splits`\n\n#### Description\nA list of indices of binary features to put at the top of each tree; ignored if `grow_policy` is `Symmetric`.\n\n**Type**\n\nlist\n\n**Default value**\n\nNone\n\n**Supported processing units**\n\nGPU\n\n## penalties\\_coefficient\nCommand-line: `--penalties-coefficient`\n\n#### Description\nA single-value common coefficient to multiply all penalties.\n\nNon-negative values are supported.\n\n**Type**\n\nfloat\n\n**Default value**\n\n1\n\n**Supported processing units**\n\nCPU\n\n## per\\_object\\_feature\\_penalties\nCommand-line: `--per-object-feature-penalties`\n\n#### Description\nPer-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time.\n\nRefer to the [Per-object and per-feature penalties](https:\/\/catboost.ai\/docs\/en\/concepts\/algorithm-score-functions) section for details on applying different score penalties.\n\nNon-negative float values are supported for each penalty.\n\nPython package\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  In this example\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n- Set the penalty individually for each required feature as an array or a dictionary (the number of features is n).\n  Format\n  ```\n  [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>]\n  ```\n  ```\n  {\"<feature index or name>\":<penalty>, .., \"<feature index or name>\":<penalty>}\n  ```\n  Array examples.\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = [0.1, 1, 3]\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = [0.1, 1, 3]\n  ```\n  These dictionary examples\n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  first_feature_use_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = {\"Feature2\":1.1,\"Feature4\":0.1}\n  ```\n  ```\n  per_object_feature_penalties = {\"2\":1.1, \"4\":0.1}\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\nR package\n\n- Set the penalty for each feature as a string (the number of features is n).\n  Format\n  ```\n  \"(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped.\n  Penalties equal to 0 at the end of the list may be dropped.\n  In this\n  example\n  \n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"(0.1,1,3)\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"(0.1,1,3)\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1.\n- Set the penalty individually for each explicitly specified feature as a string (the number of features is n).\n  Format\n  ```\n  \"<feature index or name>:<penalty>,..,<feature index or name>:<penalty>\"\n  ```\n  Note\n  \n  Spaces between values are not allowed.\n  These examples\n  \n  `first_feature_use_penalties` parameter:\n  ```\n  first_feature_use_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  first_feature_use_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  `per_object_feature_penalties` parameter:\n  ```\n  per_object_feature_penalties = \"2:1.1,4:0.1\"\n  ```\n  ```\n  per_object_feature_penalties = \"Feature2:1.1,Feature4:0.1\"\n  ```\n  are identical, given that the name of the feature indexed 2 is \"Feature2\" and the name of the feature indexed 4 is \"Feature4\".\n\n**Type**\n\n- list\n- numpy.ndarray\n- string\n- dict\n\n**Default value**\n\n0 for all objects\n\n**Supported processing units**\n\nCPU\n\n## model\\_shrink\\_rate\nCommand-line: `--model-shrink-rate`\n\n#### Description\nThe constant used to calculate the coefficient for multiplying the model on each iteration.  \n The actual model shrinkage coefficient calculated at each iteration depends on the value of the `--model-shrink-mode`for the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1\\].\n\n**Type**\n\nfloat\n\n**Default value**\n\nThe default value depends on the values of the following parameters:\n\n- `--model-shrink-mode` for the Command-line version\n- `--monotone-constraints` for the Command-line version\n\n**Supported processing units**\n\nCPU\n\n## model\\_shrink\\_mode\nCommand-line: `model_shrink_mode`\n\n#### Description\nDetermines how the actual model shrinkage coefficient is calculated at each iteration.\n\nPossible values:\n\n- Constant:\n  1 − m o d e l \\_ s h r i n k \\_ r a t e ⋅ l e a r n i n g \\_ r a t e , 1 - model\\\\\\_shrink\\\\\\_rate \\\\cdot learning\\\\\\_rate {,}\n  - m\n    \n    o\n    \n    d\n    \n    e\n    \n    l\n    \n    \\_\n    \n    s\n    \n    h\n    \n    r\n    \n    i\n    \n    n\n    \n    k\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    model\\\\\\_shrink\\\\\\_rate\n    is the value of the `--model-shrink-rate`for the Command-line version parameter.\n  - l\n    \n    e\n    \n    a\n    \n    r\n    \n    n\n    \n    i\n    \n    n\n    \n    g\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    learning\\\\\\_rate\n    is the value of the `--learning-rate`for the Command-line version parameter\n- Decreasing:\n  1 − m o d e l \\_ s h r i n k \\_ r a t e i , 1 - \\\\frac{model\\\\\\_shrink\\\\\\_rate}{i} {,}\n  - m\n    \n    o\n    \n    d\n    \n    e\n    \n    l\n    \n    \\_\n    \n    s\n    \n    h\n    \n    r\n    \n    i\n    \n    n\n    \n    k\n    \n    \\_\n    \n    r\n    \n    a\n    \n    t\n    \n    e\n    \n    model\\\\\\_shrink\\\\\\_rate\n    is the value of the `--model-shrink-rate`for the Command-line version parameter.\n  - i\n    \n    i\n    is the identifier of the iteration.\n\n**Type**\n\nstring\n\n**Default value**\n\nConstant\n\n**Supported processing units**\n\nCPU","meta_canonical":null}

3. Robots.txt Check

Query:

Response:

4. Spam/Ban Check

Query:

Response:

5. Seen Status Check

ℹ️ Skipped - page is already crawled

📄

INDEXABLE

✅

CRAWLED

1 day ago

🤖

ROBOTS ALLOWED

Page Info Filters

Filter	Status	Condition	Details
HTTP status	PASS	`download_http_code = 200`	HTTP 200
Age cutoff	PASS	`download_stamp > now() - 6 MONTH`	0.1 months ago
History drop	PASS	`isNull(history_drop_reason)`	No drop reason
Spam/ban	PASS	`fh_dont_index != 1 AND ml_spam_score = 0`	ml_spam_score=0
Canonical	PASS	`meta_canonical IS NULL OR = '' OR = src_unparsed`	Not set

Page Details

Property	Value
URL	https://catboost.ai/docs/en/references/training-parameters/common
Last Crawled	2026-04-15 22:20:28 (1 day ago)
First Indexed	2024-11-18 17:11:24 (1 year ago)
HTTP Status Code	200
Meta Title	Common parameters \| CatBoost
Meta Description	loss_function. Command-line: --loss-function. Alias: objective. Description.
Meta Canonical	null
Boilerpipe Text	loss_function Command-line: --loss-function Alias: objective Description The metric to use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the Objectives and metrics section for details on each metric). Format: <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] Supported metrics RMSE Logloss MAE CrossEntropy Quantile LogLinQuantile Lq MultiRMSE MultiClass MultiClassOneVsAll MultiLogloss MultiCrossEntropy MAPE Poisson PairLogit PairLogitPairwise QueryRMSE QuerySoftMax GroupQuantile Tweedie YetiRank YetiRankPairwise StochasticFilter StochasticRank A custom python object can also be set as the value of this parameter (see an example ). For example, use the following construction to calculate the value of Quantile with the coefficient α = 0.1 \alpha = 0.1 : Quantile:alpha=0.1 Type string object Default value Python package Depends on the class: CatBoostClassifier : Logloss if the target_border parameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass. CatBoost and CatBoostRegressor : RMSE R package, Command-line RMSE Supported processing units CPU and GPU custom_metric Command-line: --custom-metric Description Metric values to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the Objectives and metrics section for details on each metric). Format: <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] Supported metrics Examples Calculate the value of CrossEntropy: CrossEntropy Calculate the value of Quantile with the coefficient α = 0.1 \alpha = 0.1 Quantile:alpha=0.1 Calculate the values of Logloss and AUC: [ 'Logloss' , 'AUC' ] Values of all custom metrics for learn and validation datasets are saved to the Metric output files ( learn_error.tsv and test_error.tsv respectively). The directory for these files is specified in the --train-dir ( train_dir ) parameter. Use the visualization tools to see a live chart with the dynamics of the specified metrics. Type string list of strings Default value Python package None R package None Command-line None (do not output additional metric values) Supported processing units CPU and GPU eval_metric Command-line: --eval-metric Description The metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the Objectives and metrics section for details on each metric). Format: <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] Supported metrics A user-defined function can also be set as the value (see an example ). Examples: R2 Type string object Default value Optimized objective is used Supported processing units CPU and GPU iterations Command-line: -i , --iterations Aliases: num_boost_round , n_estimators , num_trees Description The maximum number of trees that can be built when solving machine learning problems. When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter. Type int Default value 1000 Supported processing units CPU and GPU learning_rate Command-line: -w , --learning-rate Alias: eta Description The learning rate. Used for reducing the gradient step. Type float Default value The default value is defined automatically for Logloss , MultiClass and RMSE loss functions depending on the number of iterations if none of parameters leaf_estimation_iterations , leaf_estimation_method , l2_leaf_reg is set. In this case, the selected learning rate is printed to stdout and saved in the model. In other cases, the default value is 0.03. Supported processing units CPU and GPU random_seed Command-line: -r , --random-seed Alias: random_state Description The random seed used for training. Type int Default value Python package None (0) R package, Command-line 0 Supported processing units CPU and GPU l2_leaf_reg Command-line: --l2-leaf-reg , l2-leaf-regularizer Alias: reg_lambda Description Coefficient at the L2 regularization term of the cost function. Any positive value is allowed. Type float Default value 3.0 Supported processing units CPU and GPU bootstrap_type Command-line: --bootstrap-type Description Bootstrap type . Defines the method for sampling the weights of objects. Supported methods: Bayesian Bernoulli MVS Poisson (supported for GPU only) No Type string Default value The default value depends on objective , task_type , bagging_temperature and sampling_unit : When the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5. Neither MultiClass nor MultiClassOneVsAll, task_type = CPU and sampling_unit = Object: MVS with the subsample parameter set to 0.8. Otherwise: Bayesian. Supported processing units CPU and GPU bagging_temperature Command-line: --bagging-temperature Description Defines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes. Use the Bayesian bootstrap to assign random weights to objects. The weights are sampled from exponential distribution if the value of this parameter is set to 1 . All weights are equal to 1 if the value of this parameter is set to 0 . Possible values are in the range [ 0 ; inf ⁡ ) [0; \inf) . The higher the value the more aggressive the bagging is. This parameter can be used if the selected bootstrap type is Bayesian. Type float Default value 1 Supported processing units CPU and GPU subsample Command-line: --subsample Description Sample rate for bagging. This parameter can be used if one of the following bootstrap types is selected: Poisson Bernoulli MVS Type float Default value The default value depends on the dataset size and the bootstrap type: Datasets with less than 100 objects — 1 Datasets with 100 objects or more: Poisson, Bernoulli — 0.66 MVS — 0.8 Supported processing units CPU and GPU sampling_frequency Command-line: --sampling-frequency Description Frequency to sample weights and objects when building trees. Supported values: PerTree — Before constructing each new tree PerTreeLevel — Before choosing each new split of a tree Type string Default value PerTreeLevel Supported processing units CPU sampling_unit Command-line: --sampling-unit Description The sampling scheme. Possible values: Object — The weight w i w_{i} of the i-th object o i o_{i} is used for sampling the corresponding object. Group — The weight w j w_{j} of the group g j g_{j} is used for sampling each object o i j o_{i_{j}} from the group g j g_{j} . Type String Default value Object Supported processing units CPU and GPU mvs_reg Command-line: --mvs-reg Description Affects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to ∞ \infty - Bernoulli). Note This parameter is supported only for the MVS sampling method (the bootstrap_type parameter must be set to MVS). Type float Default value The value is set based on the gradient distribution on the current iteration Supported processing units CPU random_strength Command-line: --random-strength Description The amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model. The value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected. The scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance. Note This parameter is not supported for the following loss functions: QueryCrossEntropy YetiRankPairwise PairLogitPairwise Type float Default value 1 Supported processing units CPU use_best_model Command-line: --use-best-model Description If this parameter is set, the number of trees that are saved in the resulting model is defined as follows: Build the number of trees defined by the training parameters. Use the validation dataset to identify the iteration with the optimal value of the metric specified in --eval-metric ( --eval-metric ). No trees are saved after this iteration. This option requires a validation dataset to be provided. Type bool Default value True if a validation set is input (the eval_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise. Supported processing units CPU and GPU best_model_min_trees Command-line: --best-model-min-trees Description The minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees. Should be used with the --use-best-model parameter. Type int Default value Python package, R package None (The minimal number of trees for the best model is not set) Command-line The minimal number of trees for the best model is not set Supported processing units CPU and GPU depth Command-line: -n , --depth Alias: max_depth Description Depth of the trees. The range of supported values depends on the processing unit type and the type of the selected loss function: CPU — Any integer up to 16. GPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions. Type int Default value 6 (16 if the growing policy is set to Lossguide) Supported processing units CPU and GPU grow_policy Command-line: --grow-policy Description The tree growing policy. Defines how to perform greedy tree construction. Possible values: SymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric. Depthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. Lossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. Type string Default value SymmetricTree Supported processing units CPU and GPU min_data_in_leaf Command-line: --min-data-in-leaf Alias: min_child_samples Description The minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value. Can be used only with the Lossguide and Depthwise growing policies. Type int Default value 1 Supported processing units CPU and GPU max_leaves Command-line: --max-leaves Alias: num_leaves Description The maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy. Note It is not recommended to use values greater than 64, since it can significantly slow down the training process. Type int Default value 31 Supported processing units CPU and GPU ignored_features Command-line: -I , --ignore-features Description Feature indices to exclude from the training. Python package It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to 42 , the corresponding non-existing feature is successfully ignored. The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to featureCount – 1 . If a file is used as input data then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: cat feature<\t>label value<\t>num feature . So for the row rock<\t>0<\t>42 , the identifier for the rock feature is 0, and for the 42 feature it's 1. For example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored: [1,2,7,42,43,44,45] R package Specifics: Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to 42 , the corresponding non-existing feature is successfully ignored. The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to featureCount – 1 . If a file is used as input data then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: cat feature<\t>label value<\t>num feature . So for the row rock<\t>0<\t>42 , the identifier for the rock feature is 0, and for the 42 feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45). Command-line It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to 42 , the corresponding non-existing feature is successfully ignored. The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to featureCount – 1 . If a file is used as input data then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: cat feature<\t>label value<\t>num feature . So for the row rock<\t>0<\t>42 , the identifier for the rock feature is 0, and for the 42 feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction: 1:2:7:42-45 . Default value Python package, R package None Command-line Omitted Supported processing units CPU and GPU one_hot_max_size Command-line: --one-hot-max-size Description Use one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features. See details . Type int Default value The default value depends on various conditions: N/A if training is performed on CPU in Pairwise scoring mode Read more about Pairwise scoring The following loss functions use Pairwise scoring: YetiRankPairwise PairLogitPairwise QueryCrossEntropy Pairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions. 255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training 10 if training is performed in Ranking mode 2 if none of the conditions above is met Supported processing units CPU and GPU has_time Command-line: --has-time Description Use the order of objects in the input data (do not perform random permutations during the Transforming categorical features to numerical features and Choosing the tree structure stages). The Timestamp column type is used to determine the order of objects if specified in the input data . Type bool Default value False (not used; generates random permutations) Supported processing units CPU and GPU rsm Command-line: --rsm Alias: colsample_bylevel Description Random subspace method. The percentage of features to use at each split selection, when features are selected over again at random. The value must be in the range (0;1]. Type float (0;1] Default value None (set to 1) Supported processing units CPU; GPU for pairwise ranking nan_mode Command-line: --nan-mode Description The method for processing missing values in the input dataset. Possible values: "Forbidden" — Missing values are not supported, their presence is interpreted as an error. "Min" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. "Max" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. Using the Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree. Type string Default value Min Supported processing units CPU and GPU input_borders Command-line: --input-borders-file Description Load Custom quantization borders and missing value modes from a file (do not generate them). Borders are automatically generated before training if this parameter is not set. Type string Default value Python package None Command-line The file is not loaded, the values are generated Supported processing units CPU and GPU output_borders Command-line: --output-borders-file Description Save quantization borders for the current dataset to a file. Refer to the file format description . Type string Default value Python package None Command-line The file is not saved Supported processing units CPU and GPU fold_permutation_block Command-line: --fold-permutation-block Description Objects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation. Type int Default value Python package 1 R package, Command-line Default value differs depending on the dataset size and ranges from 1 to 256 inclusively Supported processing units CPU and GPU leaf_estimation_method Command-line: --leaf-estimation-method Description The method used to calculate the values in leaves. Possible values: Newton Gradient Exact Type string Default value Depends on the mode and the selected loss function: Regression with Quantile or MAE loss functions — One Exact iteration. Regression with any loss function but Quantile or MAE – One Gradient iteration. Classification mode – Ten Newton iterations. Multiclassification mode – One Newton iteration. Supported processing units CPU and GPU leaf_estimation_iterations Command-line: --leaf-estimation-iterations Description CatBoost might calculate leaf values using several gradient or newton steps instead of a single one. This parameter regulates how many steps are done in every tree when calculating leaf values. Type int Default value Python package None (Depends on the training objective) R package, Command-line Depends on the training objective Supported processing units CPU and GPU leaf_estimation_backtracking Command-line: --leaf-estimation-backtracking Description When the value of the leaf_estimation_iterations parameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree. The behaviour differs depending on the value of this parameter: No — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf. Any other value —Backtracking is used. In this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf. When leaf_estimation_iterations for the Command-line version is set to n , the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having n gradient steps, the algorithm makes a single gradient step that is reduced n times, which means that it is divided by 2 ⋅ n 2\cdot n times. Possible values: No — Do not use backtracking. Supported on CPU and GPU. AnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU. Armijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU. Type string Default value AnyImprovement Supported processing units Depends on the selected value fold_len_multiplier Command-line: --fold-len-multiplier Description Coefficient for changing the length of folds. The value must be greater than 1. The best validation result is achieved with minimum values. With values close to 1 (for example, 1 + ϵ 1+\epsilon ), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects. Type float Default value 2 Supported processing units CPU and GPU approx_on_full_history Command-line: --approx-on-full-history Description The principles for calculating the approximated values. Possible values: False — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows: 1 X \frac{1}X , where X is the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate True — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate. Type bool Default value Python package, Command-line False R package True Supported processing units CPU class_weights Command-line: --class-weights Description Class weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems. Python package Note For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m _ n e g a t i v e s u m _ p o s i t i v e ) \left(\frac{sum\_negative}{sum\_positive}\right) for class 1. For example, class_weights=[0.1, 4] multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. If class labels are not standard consecutive integers [0, 1 ... class_count-1], use the dict or collections.OrderedDict type with label to weight mapping. For example, class_weights={'a': 1.0, 'b': 0.5, 'c': 2.0} multiplies the weights of objects with class label a by 1.0, the weights of objects with class label b by 0.5 and the weights of objects with class label c by 2.0. The dictionary form can also be used with standard consecutive integers class labels for additional readability. For example: class_weights={0: 1.0, 1: 0.5, 2: 2.0} . Note Class labels are extracted from dictionary keys for the following types of class_weights: dict collections.OrderedDict (when the order of classes in the model is important) The class_names parameter can be skipped when using these types. Alert Do not use this parameter with auto_class_weights and scale_pos_weight. R package For example, class_weights <- c(0.1, 4) multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. Alert Do not use this parameter with auto_class_weights. Command-line Note The quantity of class weights must match the quantity of class names specified in the --class-names parameter and the number of classes specified in the --classes-count parameter . For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m _ n e g a t i v e s u m _ p o s i t i v e ) \left(\frac{sum\_negative}{sum\_positive}\right) for class 1. Format: <value for class 1>,..,<values for class N> For example: 0.85,1.2,1 Alert Do not use this parameter with auto_class_weights. Type list dict collections.OrderedDict Default value None (the weight for all classes is set to 1) Supported processing units CPU and GPU class_names Description Classes names. Allows to redefine the default values when using the MultiClass and Logloss metrics. If the upper limit for the numeric class label is specified, the number of classes names should match this value. Warning The quantity of classes names must match the quantity of classes weights specified in the --class-weights parameter and the number of classes specified in the --classes-count parameter. Format: <name for class 1>,..,<name for class N> For example: smartphone,touchphone,tablet Type list of strings Default value None Supported processing units CPU and GPU auto_class_weights Command-line: --auto-class-weights Description Automatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights. Supported values: None — All class weights are set to 1 Balanced: C W k = m a x c = 1 K ( ∑ t i = c w i ) ∑ t i = k w i CW_k=\displaystyle\frac{max_{c=1}^K(\sum_{t_{i}=c}{w_i})}{\sum_{t_{i}=k}{w_{i}}} SqrtBalanced: C W k = m a x c = 1 K ( ∑ t i = c w i ) ∑ t i = k w i CW_k=\sqrt{\displaystyle\frac{max_{c=1}^K(\sum_{t_i=c}{w_i})}{\sum_{t_i=k}{w_i}}} Alert Do not use this parameter with class_weights and scale_pos_weight . Type string Default value None — All class weights are set to 1 Supported processing units CPU and GPU scale_pos_weight Description The weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1. Note For imbalanced datasets, the weight multiplier can be set to ( s u m _ n e g a t i v e s u m _ p o s i t i v e ) \left(\frac{sum\_negative}{sum\_positive}\right) Alert Do not use this parameter with auto_class_weights and class_weights . Type float Default value 1.0 Supported processing units CPU and GPU boosting_type Command-line: --boosting-type Description Boosting scheme. Possible values: Ordered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme. Plain — The classic gradient boosting scheme. Type string Default value Depends on the processing unit type, the number of objects in the training dataset and the selected learning mode CPU Plain GPU Any number of objects, MultiClass or MultiClassOneVsAll mode: Plain More than 50 thousand objects, any mode: Plain Less than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered Supported processing units CPU and GPU Only the Plain mode is supported for the MultiClass loss on GPU boost_from_average Command-line: --boost-from-average Description Initialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value. Available for the following loss functions: RMSE Logloss CrossEntropy Quantile MAE MAPE Type bool Default value Depends on the selected loss function: True for RMSE, Quantile, MAE, MAPE False for all other loss functions Supported processing units CPU and GPU langevin Command-line: --langevin Description Enables the Stochastic Gradient Langevin Boosting mode. Refer to the SGLB: Stochastic Gradient Langevin Boosting paper for details. Type bool Default value False Supported processing units CPU diffusion_temperature Command-line: --diffusion-temperature Description The diffusion temperature of the Stochastic Gradient Langevin Boosting mode. Only non-negative values are supported. Type float Default value 10000 Supported processing units CPU posterior_sampling Command-line: --posterior-sampling Description If this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties. Specifies options: Langevin : true, DiffusionTemperature : objects in learn pool count, ModelShrinkRate : 1 / (2. * objects in learn pool count). Type bool Default value False Supported processing units CPU only allow_const_label Command-line: --allow-const-label Description Use it to train models with datasets that have equal label values for all objects. Type bool Default value False Supported processing units CPU and GPU score_function Command-line: --score-function Description The score type used to select the next split during the tree construction. Possible values: Cosine (do not use this score type with the Lossguide tree growing policy) L2 NewtonCosine (do not use this score type with the Lossguide tree growing policy) NewtonL2 Type string Default value Cosine Supported processing units The supported score functions vary depending on the processing unit type: GPU — All score types CPU — Cosine, L2 monotone_constraints Command-line: --monotone-constraints Description Impose monotonic constraints on numerical features. Possible values: 1 — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features. -1 — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features. 0 — constraints are disabled. Supported formats for setting the value of this parameter (all feature indices are zero-based): Set constraints individually for each feature as a string (the number of features is n). Format "(<constraint_0>, <constraint_2>, .., <constraint_n-1>)" Zero constraints for features at the end of the list may be dropped. In monotone_constraints = "(1,0,-1)" an increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features. Set constraints individually for each explicitly specified feature as a string (the number of features is n). "<feature index or name>:<constraint>, .., <feature index or name>:<constraint>" These examples monotone-constraints = "2:1,4:-1" monotone-constraints = "Feature2:1,Feature4:-1" are identical, given that the name of the feature index 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Set constraints individually for each required feature as an array or a dictionary (the number of features is n). Format [<constraint_0>, <constraint_2>, .., <constraint_n-1>] {"<feature index or name>":<constraint>, .., "<feature index or name>":<constraint>} Array examples monotone_constraints = [1, 0, -1] These dictionary examples monotone_constraints = { "Feature2" : 1 , "Feature4" :- 1 } monotone_constraints = { "2" : 1 , "4" :- 1 } are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Type list of strings string dict list Default value Python package, R package None Command-line Ommited Supported processing units CPU feature_weights Command-line: --feature-weights Description Per-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split. Non-negative float values are supported for each weight. Supported formats for setting the value of this parameter: Set the multiplication weight for each feature as a string (the number of features is n). Format "(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)" Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped. In this example feature_weights = "(0.1,1,3)" the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. Set the multiplication weight individually for each explicitly specified feature as a string (the number of features is n). Format "<feature index or name>:<weight>, .., <feature index or name>:<weight>" Note Spaces between values are not allowed. These examples feature_weights = "2:1.1,4:0.1" feature_weights = "Feature2:1.1,Feature4:0.1" are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Set the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n). Format [<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>] {"<feature index or name>":<weight>, .., "<feature index or name>":<weight>} Array examples feature_weights = [0.1, 1, 3] These dictionary examples feature_weights = { "Feature2" : 1.1 , "Feature4" : 0.3 } feature_weights = { "2" : 1.1 , "4" : 0.3 } are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Type list numpy.ndarray string dict Default value 1 for all features Supported processing units CPU first_feature_use_penalties Command-line: --first-feature-use-penalties Description Per-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model. Refer to the Per-object and per-feature penalties section for details on applying different score penalties. Non-negative float values are supported for each penalty. Set the penalty for each feature as a string (the number of features is n). Format "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example first_feature_use_penalties parameter: first_feature_use_penalties = "(0.1,1,3)" per_object_feature_penalties parameter: per_object_feature_penalties = "(0.1,1,3)" Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" Note Spaces between values are not allowed. These examples first_feature_use_penalties parameter: first_feature_use_penalties = "2:1.1,4:0.1" first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" per_object_feature_penalties parameter: per_object_feature_penalties = "2:1.1,4:0.1" per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} Array examples. first_feature_use_penalties parameter: first_feature_use_penalties = [0.1, 1, 3] per_object_feature_penalties parameter: per_object_feature_penalties = [0.1, 1, 3] These dictionary examples first_feature_use_penalties parameter: first_feature_use_penalties = { "Feature2" : 1.1 , "Feature4" : 0.1 } first_feature_use_penalties = { "2" : 1.1 , "4" : 0.1 } per_object_feature_penalties parameter: per_object_feature_penalties = { "Feature2" : 1.1 , "Feature4" : 0.1 } per_object_feature_penalties = { "2" : 1.1 , "4" : 0.1 } are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Type list numpy.ndarray string dict Default value 0 for all features Supported processing units CPU fixed_binary_splits Command-line: --fixed-binary-splits Description A list of indices of binary features to put at the top of each tree; ignored if grow_policy is Symmetric . Type list Default value None Supported processing units GPU penalties_coefficient Command-line: --penalties-coefficient Description A single-value common coefficient to multiply all penalties. Non-negative values are supported. Type float Default value 1 Supported processing units CPU per_object_feature_penalties Command-line: --per-object-feature-penalties Description Per-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time. Refer to the Per-object and per-feature penalties section for details on applying different score penalties. Non-negative float values are supported for each penalty. Python package Set the penalty for each feature as a string (the number of features is n). Format "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example first_feature_use_penalties parameter: first_feature_use_penalties = "(0.1,1,3)" per_object_feature_penalties parameter: per_object_feature_penalties = "(0.1,1,3)" Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" Note Spaces between values are not allowed. These examples first_feature_use_penalties parameter: first_feature_use_penalties = "2:1.1,4:0.1" first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" per_object_feature_penalties parameter: per_object_feature_penalties = "2:1.1,4:0.1" per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} Array examples. first_feature_use_penalties parameter: first_feature_use_penalties = [0.1, 1, 3] per_object_feature_penalties parameter: per_object_feature_penalties = [0.1, 1, 3] These dictionary examples first_feature_use_penalties parameter: first_feature_use_penalties = { "Feature2" : 1.1 , "Feature4" : 0.1 } first_feature_use_penalties = { "2" : 1.1 , "4" : 0.1 } per_object_feature_penalties parameter: per_object_feature_penalties = { "Feature2" : 1.1 , "Feature4" : 0.1 } per_object_feature_penalties = { "2" : 1.1 , "4" : 0.1 } are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . R package Set the penalty for each feature as a string (the number of features is n). Format "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. Penalties equal to 0 at the end of the list may be dropped. In this example first_feature_use_penalties parameter: first_feature_use_penalties = "(0.1,1,3)" per_object_feature_penalties parameter: per_object_feature_penalties = "(0.1,1,3)" Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" Note Spaces between values are not allowed. These examples first_feature_use_penalties parameter: first_feature_use_penalties = "2:1.1,4:0.1" first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" per_object_feature_penalties parameter: per_object_feature_penalties = "2:1.1,4:0.1" per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" are identical, given that the name of the feature indexed 2 is Feature2 and the name of the feature indexed 4 is Feature4 . Type list numpy.ndarray string dict Default value 0 for all objects Supported processing units CPU model_shrink_rate Command-line: --model-shrink-rate Description The constant used to calculate the coefficient for multiplying the model on each iteration. The actual model shrinkage coefficient calculated at each iteration depends on the value of the --model-shrink-mode for the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1]. Type float Default value The default value depends on the values of the following parameters: --model-shrink-mode for the Command-line version --monotone-constraints for the Command-line version Supported processing units CPU model_shrink_mode Command-line: model_shrink_mode Description Determines how the actual model shrinkage coefficient is calculated at each iteration. Possible values: Constant: 1 − m o d e l _ s h r i n k _ r a t e ⋅ l e a r n i n g _ r a t e , 1 - model\_shrink\_rate \cdot learning\_rate {,} m o d e l _ s h r i n k _ r a t e model\_shrink\_rate is the value of the --model-shrink-rate for the Command-line version parameter. l e a r n i n g _ r a t e learning\_rate is the value of the --learning-rate for the Command-line version parameter Decreasing: 1 − m o d e l _ s h r i n k _ r a t e i , 1 - \frac{model\_shrink\_rate}{i} {,} m o d e l _ s h r i n k _ r a t e model\_shrink\_rate is the value of the --model-shrink-rate for the Command-line version parameter. i i is the identifier of the iteration. Type string Default value Constant Supported processing units CPU
Markdown	[![Logo icon](https://yastatic.net/s3/locdoc/daas-static/catboost/71b237a322eec6f2889af0dae2a9c549.svg)](https://catboost.ai/ "CatBoost") - Installation - [Overview](https://catboost.ai/docs/en/references/training-parameters/en/concepts/installation) - Python package installation - CatBoost for Apache Spark installation - R package installation - Command-line version binary - Build from source - Key Features - Training parameters - [Overview](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/) - [Common parameters](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common) - [CTR settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/ctr) - [Input settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/input) - [Multiclassification settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/multiclassification) - [Output settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/output) - [Overfitting detection settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/overfitting-detection) - [Performance settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/performance) - [Processing unit settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/processing-unit) - [Quantization settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/quantization) - [Text processing parameters](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/text-processing) - [Visualization settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/visualization) - Python package - CatBoost for Apache Spark - R package - Command-line version - Applying models - Objectives and metrics - Model analysis - Data format description - [Parameter tuning](https://catboost.ai/docs/en/references/training-parameters/en/concepts/parameter-tuning) - [Speeding up the training](https://catboost.ai/docs/en/references/training-parameters/en/concepts/speed-up-training) - Data visualization - Algorithm details - [FAQ](https://catboost.ai/docs/en/references/training-parameters/en/concepts/faq) - Educational materials - [Development and contributions](https://catboost.ai/docs/en/references/training-parameters/en/concepts/development-and-contributions) - [Contacts](https://catboost.ai/docs/en/references/training-parameters/en/concepts/contacts) loss\_function ## In this article: - [loss\_function](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#loss_function) - [custom\_metric](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#custom_metric) - [eval\_metric](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#eval_metric) - [iterations](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#iterations) - [learning\_rate](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#learning_rate) - [random\_seed](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#random_seed) - [l2\_leaf\_reg](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#l2_leaf_reg) - [bootstrap\_type](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#bootstrap_type) - [bagging\_temperature](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#bagging_temperature) - [subsample](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#subsample) - [sampling\_frequency](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#sampling_frequency) - [sampling\_unit](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#sampling_unit) - [mvs\_reg](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#mvs_reg) - [random\_strength](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#random_strength) - [use\_best\_model](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#use_best_model) - [best\_model\_min\_trees](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#best_model_min_trees) - [depth](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#depth) - [grow\_policy](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#grow_policy) - [min\_data\_in\_leaf](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#min_data_in_leaf) - [max\_leaves](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#max_leaves) - [ignored\_features](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#ignored_features) - [one\_hot\_max\_size](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#one_hot_max_size) - [has\_time](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#has_time) - [rsm](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#rsm) - [nan\_mode](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#nan_mode) - [input\_borders](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#input_borders) - [output\_borders](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#output_borders) - [fold\_permutation\_block](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fold_permutation_block) - [leaf\_estimation\_method](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_method) - [leaf\_estimation\_iterations](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_iterations) - [leaf\_estimation\_backtracking](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_backtracking) - [fold\_len\_multiplier](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fold_len_multiplier) - [approx\_on\_full\_history](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#approx_on_full_history) - [class\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#class_weights) - [class\_names](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#class_names) - [auto\_class\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#auto_class_weights) - [scale\_pos\_weight](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#scale_pos_weight) - [boosting\_type](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#boosting_type) - [boost\_from\_average](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#boost_from_average) - [langevin](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#langevin) - [diffusion\_temperature](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#diffusion_temperature) - [posterior\_sampling](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#posterior_sampling) - [allow\_const\_label](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#allow_const_label) - [score\_function](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#score_function) - [monotone\_constraints](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#monotone_constraints) - [feature\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#feature_weights) - [first\_feature\_use\_penalties](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#first_feature_use_penalties) - [fixed\_binary\_splits](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fixed_binary_splits) - [penalties\_coefficient](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#penalties_coefficient) - [per\_object\_feature\_penalties](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#per_object_feature_penalties) - [model\_shrink\_rate](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#model_shrink_rate) - [model\_shrink\_mode](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#model_shrink_mode) 1. [Training parameters](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/) 2. Common parameters # Common parameters - [loss\_function](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#loss_function) - [custom\_metric](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#custom_metric) - [eval\_metric](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#eval_metric) - [iterations](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#iterations) - [learning\_rate](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#learning_rate) - [random\_seed](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#random_seed) - [l2\_leaf\_reg](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#l2_leaf_reg) - [bootstrap\_type](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#bootstrap_type) - [bagging\_temperature](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#bagging_temperature) - [subsample](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#subsample) - [sampling\_frequency](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#sampling_frequency) - [sampling\_unit](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#sampling_unit) - [mvs\_reg](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#mvs_reg) - [random\_strength](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#random_strength) - [use\_best\_model](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#use_best_model) - [best\_model\_min\_trees](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#best_model_min_trees) - [depth](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#depth) - [grow\_policy](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#grow_policy) - [min\_data\_in\_leaf](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#min_data_in_leaf) - [max\_leaves](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#max_leaves) - [ignored\_features](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#ignored_features) - [one\_hot\_max\_size](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#one_hot_max_size) - [has\_time](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#has_time) - [rsm](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#rsm) - [nan\_mode](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#nan_mode) - [input\_borders](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#input_borders) - [output\_borders](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#output_borders) - [fold\_permutation\_block](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fold_permutation_block) - [leaf\_estimation\_method](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_method) - [leaf\_estimation\_iterations](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_iterations) - [leaf\_estimation\_backtracking](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_backtracking) - [fold\_len\_multiplier](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fold_len_multiplier) - [approx\_on\_full\_history](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#approx_on_full_history) - [class\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#class_weights) - [class\_names](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#class_names) - [auto\_class\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#auto_class_weights) - [scale\_pos\_weight](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#scale_pos_weight) - [boosting\_type](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#boosting_type) - [boost\_from\_average](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#boost_from_average) - [langevin](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#langevin) - [diffusion\_temperature](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#diffusion_temperature) - [posterior\_sampling](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#posterior_sampling) - [allow\_const\_label](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#allow_const_label) - [score\_function](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#score_function) - [monotone\_constraints](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#monotone_constraints) - [feature\_weights](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#feature_weights) - [first\_feature\_use\_penalties](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#first_feature_use_penalties) - [fixed\_binary\_splits](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#fixed_binary_splits) - [penalties\_coefficient](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#penalties_coefficient) - [per\_object\_feature\_penalties](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#per_object_feature_penalties) - [model\_shrink\_rate](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#model_shrink_rate) - [model\_shrink\_mode](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#model_shrink_mode) ## loss\_function Command-line: `--loss-function` Alias: `objective` #### Description The [metric](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions) to use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` Supported metrics - RMSE - Logloss - MAE - CrossEntropy - Quantile - LogLinQuantile - Lq - MultiRMSE - MultiClass - MultiClassOneVsAll - MultiLogloss - MultiCrossEntropy - MAPE - Poisson - PairLogit - PairLogitPairwise - QueryRMSE - QuerySoftMax - GroupQuantile - Tweedie - YetiRank - YetiRankPairwise - StochasticFilter - StochasticRank A custom python object can also be set as the value of this parameter (see an [example](https://catboost.ai/docs/en/references/training-parameters/en/concepts/python-usages-examples)). For example, use the following construction to calculate the value of Quantile with the coefficient α \= 0\.1 \\alpha = 0.1 α\=0\.1: ``` Quantile:alpha=0.1 ``` Type - string - object Default value Python package Depends on the class: - [CatBoostClassifier](https://catboost.ai/docs/en/references/training-parameters/en/concepts/python-reference_catboostclassifier): Logloss if the `target_border` parameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass. - [CatBoost](https://catboost.ai/docs/en/references/training-parameters/en/concepts/python-reference_catboost) and [CatBoostRegressor](https://catboost.ai/docs/en/references/training-parameters/en/concepts/python-reference_catboostregressor): RMSE R package, Command-line RMSE Supported processing units CPU and GPU ## custom\_metric Command-line: `--custom-metric` #### Description [Metric](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions) values to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` [Supported metrics](https://catboost.ai/docs/en/references/training-parameters/en/references/custom-metric__supported-metrics) Examples - Calculate the value of CrossEntropy: ``` CrossEntropy ``` - Calculate the value of Quantile with the coefficient α \= 0\.1 \\alpha = 0.1 α\=0\.1 ``` Quantile:alpha=0.1 ``` - Calculate the values of Logloss and AUC: ``` ['Logloss', 'AUC'] ``` Values of all custom metrics for learn and validation datasets are saved to the [Metric](https://catboost.ai/docs/en/references/training-parameters/en/concepts/output-data_loss-function) output files (`learn_error.tsv` and `test_error.tsv` respectively). The directory for these files is specified in the `--train-dir` (`train_dir`) parameter. Use the [visualization tools](https://catboost.ai/docs/en/references/training-parameters/en/features/visualization) to see a live chart with the dynamics of the specified metrics. Type - string - list of strings Default value Python package None R package None Command-line None (do not output additional metric values) Supported processing units CPU and GPU ## eval\_metric Command-line: `--eval-metric` #### Description The metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` [Supported metrics](https://catboost.ai/docs/en/references/training-parameters/en/references/eval-metric__supported-metrics) A user-defined function can also be set as the value (see an [example](https://catboost.ai/docs/en/references/training-parameters/en/concepts/python-usages-examples)). Examples: ``` R2 ``` Type - string - object Default value Optimized objective is used Supported processing units CPU and GPU ## iterations Command-line: `-i`, `--iterations` Aliases: `num_boost_round`, `n_estimators`, `num_trees` #### Description The maximum number of trees that can be built when solving machine learning problems. When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter. Type int Default value 1000 Supported processing units CPU and GPU ## learning\_rate Command-line: `-w`, `--learning-rate` Alias: `eta` #### Description The learning rate. Used for reducing the gradient step. Type float Default value The default value is defined automatically for [`Logloss`](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions-classification#Logit), [`MultiClass`](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions-multiclassification#MultiClass) and [`RMSE`](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions-regression#RMSE) loss functions depending on the number of iterations if none of parameters [`leaf_estimation_iterations`](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_iterations), [`leaf_estimation_method`](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#leaf_estimation_method), [`l2_leaf_reg`](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/common#l2_leaf_reg) is set. In this case, the selected learning rate is printed to stdout and saved in the model. In other cases, the default value is 0.03. Supported processing units CPU and GPU ## random\_seed Command-line: `-r`, `--random-seed` Alias:`random_state` #### Description The random seed used for training. Type int Default value Python package None (0) R package, Command-line 0 Supported processing units CPU and GPU ## l2\_leaf\_reg Command-line: `--l2-leaf-reg`, `l2-leaf-regularizer` Alias: `reg_lambda` #### Description Coefficient at the L2 regularization term of the cost function. Any positive value is allowed. Type float Default value 3\.0 Supported processing units CPU and GPU ## bootstrap\_type Command-line: `--bootstrap-type` #### Description [Bootstrap type](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-main-stages_bootstrap-options). Defines the method for sampling the weights of objects. Supported methods: - Bayesian - Bernoulli - MVS - Poisson (supported for GPU only) - No Type string Default value The default value depends on `objective`, `task_type`, `bagging_temperature` and `sampling_unit`: - When the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging\_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5. - Neither MultiClass nor MultiClassOneVsAll, task\_type = CPU and sampling\_unit = Object: MVS with the subsample parameter set to 0.8. - Otherwise: Bayesian. Supported processing units CPU and GPU ## bagging\_temperature Command-line: `--bagging-temperature` #### Description Defines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes. Use the Bayesian bootstrap to assign random weights to objects. The weights are sampled from exponential distribution if the value of this parameter is set to "1". All weights are equal to 1 if the value of this parameter is set to "0". Possible values are in the range \[ 0 ; inf ⁡ ) \[0; \\inf) \[0;inf). The higher the value the more aggressive the bagging is. This parameter can be used if the selected bootstrap type is Bayesian. Type float Default value 1 Supported processing units CPU and GPU ## subsample Command-line: `--subsample` #### Description Sample rate for bagging. This parameter can be used if one of the following bootstrap types is selected: - Poisson - Bernoulli - MVS Type float Default value The default value depends on the dataset size and the bootstrap type: - Datasets with less than 100 objects — 1 - Datasets with 100 objects or more: - Poisson, Bernoulli — 0.66 - MVS — 0.8 Supported processing units CPU and GPU ## sampling\_frequency Command-line: `--sampling-frequency` #### Description Frequency to sample weights and objects when building trees. Supported values: - PerTree — Before constructing each new tree - PerTreeLevel — Before choosing each new split of a tree Type string Default value PerTreeLevel Supported processing units CPU ## sampling\_unit Command-line: `--sampling-unit` #### Description The sampling scheme. Possible values: - Object — The weight w i w\_{i} wi of the i-th object o i o\_{i} oi is used for sampling the corresponding object. - Group — The weight w j w\_{j} wj of the group g j g\_{j} gj is used for sampling each object o i j o\_{i\_{j}} oij from the group g j g\_{j} gj . Type String Default value Object Supported processing units CPU and GPU ## mvs\_reg Command-line: `--mvs-reg` #### Description Affects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to ∞ \\infty ∞ - Bernoulli). Note This parameter is supported only for the MVS sampling method (the `bootstrap_type` parameter must be set to MVS). Type float Default value The value is set based on the gradient distribution on the current iteration Supported processing units CPU ## random\_strength Command-line: `--random-strength` #### Description The amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model. The value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected. The scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance. Note This parameter is not supported for the following loss functions: - QueryCrossEntropy - YetiRankPairwise - PairLogitPairwise Type float Default value 1 Supported processing units CPU ## use\_best\_model Command-line: `--use-best-model` #### Description If this parameter is set, the number of trees that are saved in the resulting model is defined as follows: 1. Build the number of trees defined by the training parameters. 2. Use the validation dataset to identify the iteration with the optimal value of the metric specified in `--eval-metric` (`--eval-metric`). No trees are saved after this iteration. This option requires a validation dataset to be provided. Type bool Default value True if a validation set is input (the eval\_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise. Supported processing units CPU and GPU ## best\_model\_min\_trees Command-line: `--best-model-min-trees` #### Description The minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees. Should be used with the `--use-best-model` parameter. Type int Default value Python package, R package None (The minimal number of trees for the best model is not set) Command-line The minimal number of trees for the best model is not set Supported processing units CPU and GPU ## depth Command-line: `-n`, `--depth` Alias: `max_depth` #### Description Depth of the trees. The range of supported values depends on the processing unit type and the type of the selected loss function: - CPU — Any integer up to 16. - GPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions. Type int Default value 6 (16 if the growing policy is set to Lossguide) Supported processing units CPU and GPU ## grow\_policy Command-line: `--grow-policy` #### Description The tree growing policy. Defines how to perform greedy tree construction. Possible values: - SymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric. - Depthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. - Lossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. Type string Default value SymmetricTree Supported processing units CPU and GPU ## min\_data\_in\_leaf Command-line: `--min-data-in-leaf` Alias: `min_child_samples` #### Description The minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value. Can be used only with the Lossguide and Depthwise growing policies. Type int Default value 1 Supported processing units CPU and GPU ## max\_leaves Command-line: `--max-leaves` Alias:`num_leaves` #### Description The maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy. Note It is not recommended to use values greater than 64, since it can significantly slow down the training process. Type int Default value 31 Supported processing units CPU and GPU ## ignored\_features Command-line: `-I`, `--ignore-features` #### Description Feature indices to exclude from the training. Python package It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored: `[1,2,7,42,43,44,45]` R package Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45). Command-line It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction: `1:2:7:42-45`. Default value Python package, R package None Command-line Omitted Supported processing units CPU and GPU ## one\_hot\_max\_size Command-line: `--one-hot-max-size` #### Description Use one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features. See [details](https://catboost.ai/docs/en/references/training-parameters/en/features/categorical-features). Type int Default value The default value depends on various conditions: - N/A if training is performed on CPU in Pairwise scoring mode Read more about Pairwise scoring The following loss functions use Pairwise scoring: - YetiRankPairwise - PairLogitPairwise - QueryCrossEntropy Pairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions. - 255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training - 10 if training is performed in [Ranking](https://catboost.ai/docs/en/references/training-parameters/en/concepts/loss-functions-ranking) mode - 2 if none of the conditions above is met Supported processing units CPU and GPU ## has\_time Command-line: `--has-time` #### Description Use the order of objects in the input data (do not perform random permutations during the [Transforming categorical features to numerical features](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-main-stages_cat-to-numberic) and [Choosing the tree structure](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-main-stages_choose-tree-structure) stages). The Timestamp column type is used to determine the order of objects if specified in the [input data](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data). Type bool Default value False (not used; generates random permutations) Supported processing units CPU and GPU ## rsm Command-line: `--rsm` Alias:`colsample_bylevel` #### Description Random subspace method. The percentage of features to use at each split selection, when features are selected over again at random. The value must be in the range (0;1\]. Type float (0;1\] Default value None (set to 1) Supported processing units CPU; GPU for pairwise ranking ## nan\_mode Command-line: `--nan-mode` #### Description The method for [processing missing values](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-missing-values-processing) in the input dataset. Possible values: - "Forbidden" — Missing values are not supported, their presence is interpreted as an error. - "Min" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. - "Max" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. Using the Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree. Note The method for processing missing values can be set individually for each feature in the [Custom quantization borders and missing value modes](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data_custom-borders) input file. Such values override the ones specified in this parameter. Type string Default value Min Supported processing units CPU and GPU ## input\_borders Command-line: `--input-borders-file` #### Description Load [Custom quantization borders and missing value modes](https://catboost.ai/docs/en/references/training-parameters/en/concepts/input-data_custom-borders) from a file (do not generate them). Borders are automatically generated before training if this parameter is not set. Type string Default value Python package None Command-line The file is not loaded, the values are generated Supported processing units CPU and GPU ## output\_borders Command-line: `--output-borders-file` #### Description Save quantization borders for the current dataset to a file. Refer to the [file format description](https://catboost.ai/docs/en/references/training-parameters/en/concepts/output-data_custom-borders). Type string Default value Python package None Command-line The file is not saved Supported processing units CPU and GPU ## fold\_permutation\_block Command-line: `--fold-permutation-block` #### Description Objects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation. Type int Default value Python package 1 R package, Command-line Default value differs depending on the dataset size and ranges from 1 to 256 inclusively Supported processing units CPU and GPU ## leaf\_estimation\_method Command-line: `--leaf-estimation-method` #### Description The method used to calculate the values in leaves. Possible values: - Newton - Gradient - Exact Type string Default value Depends on the mode and the selected loss function: - Regression with Quantile or MAE loss functions — One Exact iteration. - Regression with any loss function but Quantile or MAE – One Gradient iteration. - Classification mode – Ten Newton iterations. - Multiclassification mode – One Newton iteration. Supported processing units CPU and GPU ## leaf\_estimation\_iterations Command-line: `--leaf-estimation-iterations` #### Description CatBoost might calculate leaf values using several gradient or newton steps instead of a single one. This parameter regulates how many steps are done in every tree when calculating leaf values. Type int Default value Python package None (Depends on the training objective) R package, Command-line Depends on the training objective Supported processing units CPU and GPU ## leaf\_estimation\_backtracking Command-line: `--leaf-estimation-backtracking` #### Description When the value of the `leaf_estimation_iterations` parameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree. The behaviour differs depending on the value of this parameter: - No — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf. - Any other value —Backtracking is used. In this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf. When `leaf_estimation_iterations` for the Command-line version is set to `n`, the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having `n` gradient steps, the algorithm makes a single gradient step that is reduced `n` times, which means that it is divided by 2 ⋅ n 2\\cdot n 2⋅n times. Possible values: - No — Do not use backtracking. Supported on CPU and GPU. - AnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU. - Armijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU. Type string Default value AnyImprovement Supported processing units Depends on the selected value ## fold\_len\_multiplier Command-line: `--fold-len-multiplier` #### Description Coefficient for changing the length of folds. The value must be greater than 1. The best validation result is achieved with minimum values. With values close to 1 (for example, 1 \+ ϵ 1+\\epsilon 1\+ϵ), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects. Type float Default value 2 Supported processing units CPU and GPU ## approx\_on\_full\_history Command-line:`--approx-on-full-history` #### Description The principles for calculating the approximated values. Possible values: - "False" — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows: 1 X \\frac{1}X X1 , where `X` is the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate - "True" — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate. Type bool Default value Python package, Command-line False R package True Supported processing units CPU ## class\_weights Command-line: `--class-weights` #### Description Class weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems. Python package Note For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) (sum\_positivesum\_negative) for class 1. For example, `class_weights=[0.1, 4]`multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. If class labels are not standard consecutive integers \[0, 1 ... class\_count-1\], use the dict or collections.OrderedDict type with label to weight mapping. For example, `class_weights={'a': 1.0, 'b': 0.5, 'c': 2.0}` multiplies the weights of objects with class label `a` by 1.0, the weights of objects with class label `b` by 0.5 and the weights of objects with class label `c` by 2.0. The dictionary form can also be used with standard consecutive integers class labels for additional readability. For example: `class_weights={0: 1.0, 1: 0.5, 2: 2.0}`. Note Class labels are extracted from dictionary keys for the following types of class\_weights: - dict - collections.OrderedDict (when the order of classes in the model is important) The class\_names parameter can be skipped when using these types. Alert Do not use this parameter with auto\_class\_weights and scale\_pos\_weight. R package For example, `class_weights <- c(0.1, 4)` multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. Alert Do not use this parameter with auto\_class\_weights. Command-line Note The quantity of class weights must match the quantity of class names specified in the `--class-names` parameter and the number of classes specified in the `--classes-count parameter`. For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) (sum\_positivesum\_negative) for class 1. Format: ``` <value for class 1>,..,<values for class N> ``` For example: ``` 0.85,1.2,1 ``` Alert Do not use this parameter with auto\_class\_weights. Type - list - dict - collections.OrderedDict Default value None (the weight for all classes is set to 1) Supported processing units CPU and GPU ## class\_names #### Description Classes names. Allows to redefine the default values when using the MultiClass and Logloss metrics. If the upper limit for the numeric class label is specified, the number of classes names should match this value. Warning The quantity of classes names must match the quantity of classes weights specified in the `--class-weights` parameter and the number of classes specified in the `--classes-count` parameter. Format: ``` <name for class 1>,..,<name for class N> ``` For example: ``` smartphone,touchphone,tablet ``` Type list of strings Default value None Supported processing units CPU and GPU ## auto\_class\_weights Command-line: `--auto-class-weights` #### Description Automatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights. Supported values: - None — All class weights are set to 1 - Balanced: C W k \= m a x c \= 1 K ( ∑ t i \= c w i ) ∑ t i \= k w i CW\_k=\\displaystyle\\frac{max\_{c=1}^K(\\sum\_{t\_{i}=c}{w\_i})}{\\sum\_{t\_{i}=k}{w\_{i}}} CWk\=∑ti\=kwimaxc\=1K(∑ti\=cwi) - SqrtBalanced: C W k \= m a x c \= 1 K ( ∑ t i \= c w i ) ∑ t i \= k w i CW\_k=\\sqrt{\\displaystyle\\frac{max\_{c=1}^K(\\sum\_{t\_i=c}{w\_i})}{\\sum\_{t\_i=k}{w\_i}}} CWk\= ∑ti\=kwimaxc\=1K(∑ti\=cwi) Alert Do not use this parameter with `class_weights` and `scale_pos_weight`. Type string Default value None — All class weights are set to 1 Supported processing units CPU and GPU ## scale\_pos\_weight #### Description The weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1. Note For imbalanced datasets, the weight multiplier can be set to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) (sum\_positivesum\_negative) Alert Do not use this parameter with `auto_class_weights` and `class_weights`. Type float Default value 1\.0 Supported processing units CPU and GPU ## boosting\_type Command-line: `--boosting-type` #### Description Boosting scheme. Possible values: - Ordered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme. - Plain — The classic gradient boosting scheme. Type string Default value Depends on the processing unit type, the number of objects in the training dataset and the selected learning mode - CPU Plain - GPU - Any number of objects, MultiClass or MultiClassOneVsAll mode: Plain - More than 50 thousand objects, any mode: Plain - Less than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered Supported processing units CPU and GPU Only the Plain mode is supported for the MultiClass loss on GPU ## boost\_from\_average Command-line: `--boost-from-average` #### Description Initialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value. Available for the following loss functions: - RMSE - Logloss - CrossEntropy - Quantile - MAE - MAPE Type bool Default value Depends on the selected loss function: - True for RMSE, Quantile, MAE, MAPE - False for all other loss functions Supported processing units CPU and GPU ## langevin Command-line: `--langevin` #### Description Enables the Stochastic Gradient Langevin Boosting mode. Refer to the [SGLB: Stochastic Gradient Langevin Boosting](https://arxiv.org/abs/2001.07248) paper for details. Type bool Default value False Supported processing units CPU ## diffusion\_temperature Command-line: `--diffusion-temperature` #### Description The diffusion temperature of the Stochastic Gradient Langevin Boosting mode. Only non-negative values are supported. Type float Default value 10000 Supported processing units CPU ## posterior\_sampling Command-line: `--posterior-sampling` #### Description If this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties. Specifies options: - `Langevin`: true, - `DiffusionTemperature`: objects in learn pool count, - `ModelShrinkRate`: 1 / (2. \* objects in learn pool count). Type bool Default value False Supported processing units CPU only ## allow\_const\_label Command-line: `--allow-const-label` #### Description Use it to train models with datasets that have equal label values for all objects. Type bool Default value False Supported processing units CPU and GPU ## score\_function Command-line: `--score-function` #### Description The [score type](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-score-functions) used to select the next split during the tree construction. Possible values: - Cosine (do not use this score type with the Lossguide tree growing policy) - L2 - NewtonCosine (do not use this score type with the Lossguide tree growing policy) - NewtonL2 Type string Default value Cosine Supported processing units The supported score functions vary depending on the processing unit type: - GPU — All score types - CPU — Cosine, L2 ## monotone\_constraints Command-line: `--monotone-constraints` #### Description Impose monotonic constraints on numerical features. Possible values: - "1" — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features. - "\-1" — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features. - "0" — constraints are disabled. Supported formats for setting the value of this parameter (all feature indices are zero-based): - Set constraints individually for each feature as a string (the number of features is n). Format ``` "(<constraint_0>, <constraint_2>, .., <constraint_n-1>)" ``` Zero constraints for features at the end of the list may be dropped. In `monotone_constraints = "(1,0,-1)"`an increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features. - Set constraints individually for each explicitly specified feature as a string (the number of features is n). ``` "<feature index or name>:<constraint>, .., <feature index or name>:<constraint>" ``` These examples ``` monotone-constraints = "2:1,4:-1" ``` ``` monotone-constraints = "Feature2:1,Feature4:-1" ``` are identical, given that the name of the feature index 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set constraints individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<constraint_0>, <constraint_2>, .., <constraint_n-1>] ``` ``` {"<feature index or name>":<constraint>, .., "<feature index or name>":<constraint>} ``` Array examples ``` monotone_constraints = [1, 0, -1] ``` These dictionary examples ``` monotone_constraints = {"Feature2":1,"Feature4":-1} ``` ``` monotone_constraints = {"2":1, "4":-1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list of strings - string - dict - list Default value Python package, R package None Command-line Ommited Supported processing units CPU ## feature\_weights Command-line: `--feature-weights` #### Description Per-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split. Non-negative float values are supported for each weight. Supported formats for setting the value of this parameter: - Set the multiplication weight for each feature as a string (the number of features is n). Format ``` "(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped. In this example ``` feature_weights = "(0.1,1,3)" ``` the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the multiplication weight individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<weight>, .., <feature index or name>:<weight>" ``` Note Spaces between values are not allowed. These examples ``` feature_weights = "2:1.1,4:0.1" ``` ``` feature_weights = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>] ``` ``` {"<feature index or name>":<weight>, .., "<feature index or name>":<weight>} ``` Array examples ``` feature_weights = [0.1, 1, 3] ``` These dictionary examples ``` feature_weights = {"Feature2":1.1,"Feature4":0.3} ``` ``` feature_weights = {"2":1.1, "4":0.3} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 1 for all features Supported processing units CPU ## first\_feature\_use\_penalties Command-line: `--first-feature-use-penalties` #### Description Per-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model. Refer to the [Per-object and per-feature penalties](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-score-functions) section for details on applying different score penalties. Non-negative float values are supported for each penalty. - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] ``` ``` {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} ``` Array examples. `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = [0.1, 1, 3] ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = [0.1, 1, 3] ``` These dictionary examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` first_feature_use_penalties = {"2":1.1, "4":0.1} ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` per_object_feature_penalties = {"2":1.1, "4":0.1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 0 for all features Supported processing units CPU ## fixed\_binary\_splits Command-line: `--fixed-binary-splits` #### Description A list of indices of binary features to put at the top of each tree; ignored if `grow_policy` is `Symmetric`. Type list Default value None Supported processing units GPU ## penalties\_coefficient Command-line: `--penalties-coefficient` #### Description A single-value common coefficient to multiply all penalties. Non-negative values are supported. Type float Default value 1 Supported processing units CPU ## per\_object\_feature\_penalties Command-line: `--per-object-feature-penalties` #### Description Per-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time. Refer to the [Per-object and per-feature penalties](https://catboost.ai/docs/en/references/training-parameters/en/concepts/algorithm-score-functions) section for details on applying different score penalties. Non-negative float values are supported for each penalty. Python package - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] ``` ``` {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} ``` Array examples. `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = [0.1, 1, 3] ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = [0.1, 1, 3] ``` These dictionary examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` first_feature_use_penalties = {"2":1.1, "4":0.1} ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` per_object_feature_penalties = {"2":1.1, "4":0.1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". R package - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 0 for all objects Supported processing units CPU ## model\_shrink\_rate Command-line: `--model-shrink-rate` #### Description The constant used to calculate the coefficient for multiplying the model on each iteration. The actual model shrinkage coefficient calculated at each iteration depends on the value of the `--model-shrink-mode`for the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1\]. Type float Default value The default value depends on the values of the following parameters: - `--model-shrink-mode` for the Command-line version - `--monotone-constraints` for the Command-line version Supported processing units CPU ## model\_shrink\_mode Command-line: `model_shrink_mode` #### Description Determines how the actual model shrinkage coefficient is calculated at each iteration. Possible values: - Constant: 1 − m o d e l \_ s h r i n k \_ r a t e ⋅ l e a r n i n g \_ r a t e , 1 - model\\\_shrink\\\_rate \\cdot learning\\\_rate {,} 1−model\_shrink\_rate⋅learning\_rate, - m o d e l \_ s h r i n k \_ r a t e model\\\_shrink\\\_rate model\_shrink\_rate is the value of the `--model-shrink-rate`for the Command-line version parameter. - l e a r n i n g \_ r a t e learning\\\_rate learning\_rate is the value of the `--learning-rate`for the Command-line version parameter - Decreasing: 1 − m o d e l \_ s h r i n k \_ r a t e i , 1 - \\frac{model\\\_shrink\\\_rate}{i} {,} 1−imodel\_shrink\_rate, - m o d e l \_ s h r i n k \_ r a t e model\\\_shrink\\\_rate model\_shrink\_rate is the value of the `--model-shrink-rate`for the Command-line version parameter. - i i i is the identifier of the iteration. Type string Default value Constant Supported processing units CPU ### Was the article helpful? Yes No Previous [Overview](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/) Next [CTR settings](https://catboost.ai/docs/en/references/training-parameters/en/references/training-parameters/ctr) ![](https://mc.yandex.ru/watch/60763294)
Readable Markdown	## loss\_function Command-line: `--loss-function` Alias: `objective` #### Description The [metric](https://catboost.ai/docs/en/concepts/loss-functions) to use in training. The specified value also determines the machine learning problem to solve. Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` Supported metrics - RMSE - Logloss - MAE - CrossEntropy - Quantile - LogLinQuantile - Lq - MultiRMSE - MultiClass - MultiClassOneVsAll - MultiLogloss - MultiCrossEntropy - MAPE - Poisson - PairLogit - PairLogitPairwise - QueryRMSE - QuerySoftMax - GroupQuantile - Tweedie - YetiRank - YetiRankPairwise - StochasticFilter - StochasticRank A custom python object can also be set as the value of this parameter (see an [example](https://catboost.ai/docs/en/concepts/python-usages-examples)). For example, use the following construction to calculate the value of Quantile with the coefficient α \= 0\.1 \\alpha = 0.1: ``` Quantile:alpha=0.1 ``` Type - string - object Default value Python package Depends on the class: - [CatBoostClassifier](https://catboost.ai/docs/en/concepts/python-reference_catboostclassifier): Logloss if the `target_border` parameter value differs from None. Otherwise, the default loss function depends on the number of unique target values and is either set to Logloss or MultiClass. - [CatBoost](https://catboost.ai/docs/en/concepts/python-reference_catboost) and [CatBoostRegressor](https://catboost.ai/docs/en/concepts/python-reference_catboostregressor): RMSE R package, Command-line RMSE Supported processing units CPU and GPU ## custom\_metric Command-line: `--custom-metric` #### Description [Metric](https://catboost.ai/docs/en/concepts/loss-functions) values to output during training. These functions are not optimized and are displayed for informational purposes only. Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` [Supported metrics](https://catboost.ai/docs/en/references/custom-metric__supported-metrics) Examples - Calculate the value of CrossEntropy: ``` CrossEntropy ``` - Calculate the value of Quantile with the coefficient α \= 0\.1 \\alpha = 0.1 ``` Quantile:alpha=0.1 ``` - Calculate the values of Logloss and AUC: ``` ['Logloss', 'AUC'] ``` Values of all custom metrics for learn and validation datasets are saved to the [Metric](https://catboost.ai/docs/en/concepts/output-data_loss-function) output files (`learn_error.tsv` and `test_error.tsv` respectively). The directory for these files is specified in the `--train-dir` (`train_dir`) parameter. Use the [visualization tools](https://catboost.ai/docs/en/features/visualization) to see a live chart with the dynamics of the specified metrics. Type - string - list of strings Default value Python package None R package None Command-line None (do not output additional metric values) Supported processing units CPU and GPU ## eval\_metric Command-line: `--eval-metric` #### Description The metric used for overfitting detection (if enabled) and best model selection (if enabled). Some metrics support optional parameters (see the [Objectives and metrics](https://catboost.ai/docs/en/concepts/loss-functions) section for details on each metric). Format: ``` <Metric>[:<parameter 1>=<value>;..;<parameter N>=<value>] ``` [Supported metrics](https://catboost.ai/docs/en/references/eval-metric__supported-metrics) A user-defined function can also be set as the value (see an [example](https://catboost.ai/docs/en/concepts/python-usages-examples)). Examples: ``` R2 ``` Type - string - object Default value Optimized objective is used Supported processing units CPU and GPU ## iterations Command-line: `-i`, `--iterations` Aliases: `num_boost_round`, `n_estimators`, `num_trees` #### Description The maximum number of trees that can be built when solving machine learning problems. When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter. Type int Default value 1000 Supported processing units CPU and GPU ## learning\_rate Command-line: `-w`, `--learning-rate` Alias: `eta` #### Description The learning rate. Used for reducing the gradient step. Type float Default value The default value is defined automatically for [`Logloss`](https://catboost.ai/docs/en/concepts/loss-functions-classification#Logit), [`MultiClass`](https://catboost.ai/docs/en/concepts/loss-functions-multiclassification#MultiClass) and [`RMSE`](https://catboost.ai/docs/en/concepts/loss-functions-regression#RMSE) loss functions depending on the number of iterations if none of parameters [`leaf_estimation_iterations`](https://catboost.ai/docs/en/references/training-parameters/common#leaf_estimation_iterations), [`leaf_estimation_method`](https://catboost.ai/docs/en/references/training-parameters/common#leaf_estimation_method), [`l2_leaf_reg`](https://catboost.ai/docs/en/references/training-parameters/common#l2_leaf_reg) is set. In this case, the selected learning rate is printed to stdout and saved in the model. In other cases, the default value is 0.03. Supported processing units CPU and GPU ## random\_seed Command-line: `-r`, `--random-seed` Alias:`random_state` #### Description The random seed used for training. Type int Default value Python package None (0) R package, Command-line 0 Supported processing units CPU and GPU ## l2\_leaf\_reg Command-line: `--l2-leaf-reg`, `l2-leaf-regularizer` Alias: `reg_lambda` #### Description Coefficient at the L2 regularization term of the cost function. Any positive value is allowed. Type float Default value 3\.0 Supported processing units CPU and GPU ## bootstrap\_type Command-line: `--bootstrap-type` #### Description [Bootstrap type](https://catboost.ai/docs/en/concepts/algorithm-main-stages_bootstrap-options). Defines the method for sampling the weights of objects. Supported methods: - Bayesian - Bernoulli - MVS - Poisson (supported for GPU only) - No Type string Default value The default value depends on `objective`, `task_type`, `bagging_temperature` and `sampling_unit`: - When the objective parameter is QueryCrossEntropy, YetiRankPairwise, PairLogitPairwise and the bagging\_temperature parameter is not set: Bernoulli with the subsample parameter set to 0.5. - Neither MultiClass nor MultiClassOneVsAll, task\_type = CPU and sampling\_unit = Object: MVS with the subsample parameter set to 0.8. - Otherwise: Bayesian. Supported processing units CPU and GPU ## bagging\_temperature Command-line: `--bagging-temperature` #### Description Defines the settings of the Bayesian bootstrap. It is used by default in classification and regression modes. Use the Bayesian bootstrap to assign random weights to objects. The weights are sampled from exponential distribution if the value of this parameter is set to "1". All weights are equal to 1 if the value of this parameter is set to "0". Possible values are in the range \[ 0 ; inf ⁡ ) \[0; \\inf). The higher the value the more aggressive the bagging is. This parameter can be used if the selected bootstrap type is Bayesian. Type float Default value 1 Supported processing units CPU and GPU ## subsample Command-line: `--subsample` #### Description Sample rate for bagging. This parameter can be used if one of the following bootstrap types is selected: - Poisson - Bernoulli - MVS Type float Default value The default value depends on the dataset size and the bootstrap type: - Datasets with less than 100 objects — 1 - Datasets with 100 objects or more: - Poisson, Bernoulli — 0.66 - MVS — 0.8 Supported processing units CPU and GPU ## sampling\_frequency Command-line: `--sampling-frequency` #### Description Frequency to sample weights and objects when building trees. Supported values: - PerTree — Before constructing each new tree - PerTreeLevel — Before choosing each new split of a tree Type string Default value PerTreeLevel Supported processing units CPU ## sampling\_unit Command-line: `--sampling-unit` #### Description The sampling scheme. Possible values: - Object — The weight w i w\_{i} of the i-th object o i o\_{i} is used for sampling the corresponding object. - Group — The weight w j w\_{j} of the group g j g\_{j} is used for sampling each object o i j o\_{i\_{j}} from the group g j g\_{j} . Type String Default value Object Supported processing units CPU and GPU ## mvs\_reg Command-line: `--mvs-reg` #### Description Affects the weight of the denominator and can be used for balancing between the importance and Bernoulli sampling (setting it to 0 implies importance sampling and to ∞ \\infty - Bernoulli). Note This parameter is supported only for the MVS sampling method (the `bootstrap_type` parameter must be set to MVS). Type float Default value The value is set based on the gradient distribution on the current iteration Supported processing units CPU ## random\_strength Command-line: `--random-strength` #### Description The amount of randomness to use for scoring splits when the tree structure is selected. Use this parameter to avoid overfitting the model. The value of this parameter is used when selecting splits. On every iteration each possible split gets a score (for example, the score indicates how much adding this split will improve the loss function for the training dataset). The split with the highest score is selected. The scores have no randomness. A normally distributed random variable is added to the score of the feature. It has a zero mean and a variance that decreases during the training. The value of this parameter is the multiplier of the variance. Note This parameter is not supported for the following loss functions: - QueryCrossEntropy - YetiRankPairwise - PairLogitPairwise Type float Default value 1 Supported processing units CPU ## use\_best\_model Command-line: `--use-best-model` #### Description If this parameter is set, the number of trees that are saved in the resulting model is defined as follows: 1. Build the number of trees defined by the training parameters. 2. Use the validation dataset to identify the iteration with the optimal value of the metric specified in `--eval-metric` (`--eval-metric`). No trees are saved after this iteration. This option requires a validation dataset to be provided. Type bool Default value True if a validation set is input (the eval\_set parameter is defined) and at least one of the label values of objects in this set differs from the others. False otherwise. Supported processing units CPU and GPU ## best\_model\_min\_trees Command-line: `--best-model-min-trees` #### Description The minimal number of trees that the best model should have. If set, the output model contains at least the given number of trees even if the optimal value of the evaluation metric on the validation dataset is achieved with smaller number of trees. Should be used with the `--use-best-model` parameter. Type int Default value Python package, R package None (The minimal number of trees for the best model is not set) Command-line The minimal number of trees for the best model is not set Supported processing units CPU and GPU ## depth Command-line: `-n`, `--depth` Alias: `max_depth` #### Description Depth of the trees. The range of supported values depends on the processing unit type and the type of the selected loss function: - CPU — Any integer up to 16. - GPU — Any integer up to 8 for pairwise modes (YetiRank, PairLogitPairwise, and QueryCrossEntropy), and up to 16 for all other loss functions. Type int Default value 6 (16 if the growing policy is set to Lossguide) Supported processing units CPU and GPU ## grow\_policy Command-line: `--grow-policy` #### Description The tree growing policy. Defines how to perform greedy tree construction. Possible values: - SymmetricTree —A tree is built level by level until the specified depth is reached. On each iteration, all leaves from the last tree level are split with the same condition. The resulting tree structure is always symmetric. - Depthwise — A tree is built level by level until the specified depth is reached. On each iteration, all non-terminal leaves from the last tree level are split. Each leaf is split by condition with the best loss improvement. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. - Lossguide — A tree is built leaf by leaf until the specified maximum number of leaves is reached. On each iteration, non-terminal leaf with the best loss improvement is split. Note Models with this growing policy can not be analyzed using the PredictionDiff feature importance and can be exported only to json and cbm. Type string Default value SymmetricTree Supported processing units CPU and GPU ## min\_data\_in\_leaf Command-line: `--min-data-in-leaf` Alias: `min_child_samples` #### Description The minimum number of training samples in a leaf. CatBoost does not search for new splits in leaves with samples count less than the specified value. Can be used only with the Lossguide and Depthwise growing policies. Type int Default value 1 Supported processing units CPU and GPU ## max\_leaves Command-line: `--max-leaves` Alias:`num_leaves` #### Description The maximum number of leafs in the resulting tree. Can be used only with the Lossguide growing policy. Note It is not recommended to use values greater than 64, since it can significantly slow down the training process. Type int Default value 31 Supported processing units CPU and GPU ## ignored\_features Command-line: `-I`, `--ignore-features` #### Description Feature indices to exclude from the training. Python package It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, use the following construction if features indexed 1, 2, 7, 42, 43, 44, 45, should be ignored: `[1,2,7,42,43,44,45]` R package Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, the value of this parameter should be set to c(1,2,7,42,43,44,45). Command-line It is assumed that all passed values are feature names if at least one of the passed values can not be converted to a number or a range of numbers. Otherwise, it is assumed that all passed values are feature indices. Specifics: - Non-negative indices that do not match any features are successfully ignored. For example, if five features are defined for the objects in the dataset and this parameter is set to "42", the corresponding non-existing feature is successfully ignored. - The identifier corresponds to the feature's index. Feature indices used in train and feature importance are numbered from 0 to `featureCount – 1`. If a file is used as [input data](https://catboost.ai/docs/en/concepts/input-data) then any non-feature column types are ignored when calculating these indices. For example, each row in the input file contains data in the following order: `cat feature<\t>label value<\t>num feature`. So for the row `rock<\t>0<\t>42`, the identifier for the "rock" feature is 0, and for the "42" feature it's 1. For example, if training should exclude features with the identifiers 1, 2, 7, 42, 43, 44, 45, use the following construction: `1:2:7:42-45`. Default value Python package, R package None Command-line Omitted Supported processing units CPU and GPU ## one\_hot\_max\_size Command-line: `--one-hot-max-size` #### Description Use one-hot encoding for all categorical features with a number of different values less than or equal to the given parameter value. Ctrs are not calculated for such features. See [details](https://catboost.ai/docs/en/features/categorical-features). Type int Default value The default value depends on various conditions: - N/A if training is performed on CPU in Pairwise scoring mode Read more about Pairwise scoring The following loss functions use Pairwise scoring: - YetiRankPairwise - PairLogitPairwise - QueryCrossEntropy Pairwise scoring is slightly different from regular training on pairs, since pairs are generated only internally during the training for the corresponding metrics. One-hot encoding is not available for these loss functions. - 255 if training is performed on GPU and the selected Ctr types require target data that is not available during the training - 10 if training is performed in [Ranking](https://catboost.ai/docs/en/concepts/loss-functions-ranking) mode - 2 if none of the conditions above is met Supported processing units CPU and GPU ## has\_time Command-line: `--has-time` #### Description Use the order of objects in the input data (do not perform random permutations during the [Transforming categorical features to numerical features](https://catboost.ai/docs/en/concepts/algorithm-main-stages_cat-to-numberic) and [Choosing the tree structure](https://catboost.ai/docs/en/concepts/algorithm-main-stages_choose-tree-structure) stages). The Timestamp column type is used to determine the order of objects if specified in the [input data](https://catboost.ai/docs/en/concepts/input-data). Type bool Default value False (not used; generates random permutations) Supported processing units CPU and GPU ## rsm Command-line: `--rsm` Alias:`colsample_bylevel` #### Description Random subspace method. The percentage of features to use at each split selection, when features are selected over again at random. The value must be in the range (0;1\]. Type float (0;1\] Default value None (set to 1) Supported processing units CPU; GPU for pairwise ranking ## nan\_mode Command-line: `--nan-mode` #### Description The method for [processing missing values](https://catboost.ai/docs/en/concepts/algorithm-missing-values-processing) in the input dataset. Possible values: - "Forbidden" — Missing values are not supported, their presence is interpreted as an error. - "Min" — Missing values are processed as the minimum value (less than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. - "Max" — Missing values are processed as the maximum value (greater than all other values) for the feature. It is guaranteed that a split that separates missing values from all other values is considered when selecting trees. Using the Min or Max value of this parameter guarantees that a split between missing values and other values is considered when selecting a new split in the tree. Type string Default value Min Supported processing units CPU and GPU ## input\_borders Command-line: `--input-borders-file` #### Description Load [Custom quantization borders and missing value modes](https://catboost.ai/docs/en/concepts/input-data_custom-borders) from a file (do not generate them). Borders are automatically generated before training if this parameter is not set. Type string Default value Python package None Command-line The file is not loaded, the values are generated Supported processing units CPU and GPU ## output\_borders Command-line: `--output-borders-file` #### Description Save quantization borders for the current dataset to a file. Refer to the [file format description](https://catboost.ai/docs/en/concepts/output-data_custom-borders). Type string Default value Python package None Command-line The file is not saved Supported processing units CPU and GPU ## fold\_permutation\_block Command-line: `--fold-permutation-block` #### Description Objects in the dataset are grouped in blocks before the random permutations. This parameter defines the size of the blocks. The smaller is the value, the slower is the training. Large values may result in quality degradation. Type int Default value Python package 1 R package, Command-line Default value differs depending on the dataset size and ranges from 1 to 256 inclusively Supported processing units CPU and GPU ## leaf\_estimation\_method Command-line: `--leaf-estimation-method` #### Description The method used to calculate the values in leaves. Possible values: - Newton - Gradient - Exact Type string Default value Depends on the mode and the selected loss function: - Regression with Quantile or MAE loss functions — One Exact iteration. - Regression with any loss function but Quantile or MAE – One Gradient iteration. - Classification mode – Ten Newton iterations. - Multiclassification mode – One Newton iteration. Supported processing units CPU and GPU ## leaf\_estimation\_iterations Command-line: `--leaf-estimation-iterations` #### Description CatBoost might calculate leaf values using several gradient or newton steps instead of a single one. This parameter regulates how many steps are done in every tree when calculating leaf values. Type int Default value Python package None (Depends on the training objective) R package, Command-line Depends on the training objective Supported processing units CPU and GPU ## leaf\_estimation\_backtracking Command-line: `--leaf-estimation-backtracking` #### Description When the value of the `leaf_estimation_iterations` parameter is greater than 1, CatBoost makes several gradient or newton steps when calculating the resulting leaf values of a tree. The behaviour differs depending on the value of this parameter: - No — Every next step is a regular gradient or newton step: the gradient step is calculated and added to the leaf. - Any other value —Backtracking is used. In this case, before adding a step, a condition is checked. If the condition is not met, then the step size is reduced (divided by 2), otherwise the step is added to the leaf. When `leaf_estimation_iterations` for the Command-line version is set to `n`, the leaf estimation iterations are calculated as follows: each iteration is either an addition of the next step to the leaf value, or it's a scaling of the leaf value. Scaling counts as a separate iteration. Thus, it is possible that instead of having `n` gradient steps, the algorithm makes a single gradient step that is reduced `n` times, which means that it is divided by 2 ⋅ n 2\\cdot n times. Possible values: - No — Do not use backtracking. Supported on CPU and GPU. - AnyImprovement — Reduce the descent step up to the point when the loss function value is smaller than it was on the previous step. The trial reduction factors are 2, 4, 8, and so on. Supported on CPU and GPU. - Armijo — Reduce the descent step until the Armijo condition is met. Supported only on GPU. Type string Default value AnyImprovement Supported processing units Depends on the selected value ## fold\_len\_multiplier Command-line: `--fold-len-multiplier` #### Description Coefficient for changing the length of folds. The value must be greater than 1. The best validation result is achieved with minimum values. With values close to 1 (for example, 1 \+ ϵ 1+\\epsilon), each iteration takes a quadratic amount of memory and time for the number of objects in the iteration. Thus, low values are possible only when there is a small number of objects. Type float Default value 2 Supported processing units CPU and GPU ## approx\_on\_full\_history Command-line:`--approx-on-full-history` #### Description The principles for calculating the approximated values. Possible values: - "False" — Use only а fraction of the fold for calculating the approximated values. The size of the fraction is calculated as follows: 1 X \\frac{1}X , where `X` is the specified coefficient for changing the length of folds. This mode is faster and in rare cases slightly less accurate - "True" — Use all the preceding rows in the fold for calculating the approximated values. This mode is slower and in rare cases slightly more accurate. Type bool Default value Python package, Command-line False R package True Supported processing units CPU ## class\_weights Command-line: `--class-weights` #### Description Class weights. The values are used as multipliers for the object weights. This parameter can be used for solving binary classification and multiclassification problems. Python package Note For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) for class 1. For example, `class_weights=[0.1, 4]`multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. If class labels are not standard consecutive integers \[0, 1 ... class\_count-1\], use the dict or collections.OrderedDict type with label to weight mapping. For example, `class_weights={'a': 1.0, 'b': 0.5, 'c': 2.0}` multiplies the weights of objects with class label `a` by 1.0, the weights of objects with class label `b` by 0.5 and the weights of objects with class label `c` by 2.0. The dictionary form can also be used with standard consecutive integers class labels for additional readability. For example: `class_weights={0: 1.0, 1: 0.5, 2: 2.0}`. Note Class labels are extracted from dictionary keys for the following types of class\_weights: - dict - collections.OrderedDict (when the order of classes in the model is important) The class\_names parameter can be skipped when using these types. Alert Do not use this parameter with auto\_class\_weights and scale\_pos\_weight. R package For example, `class_weights <- c(0.1, 4)` multiplies the weights of objects from class 0 by 0.1 and the weights of objects from class 1 by 4. Alert Do not use this parameter with auto\_class\_weights. Command-line Note The quantity of class weights must match the quantity of class names specified in the `--class-names` parameter and the number of classes specified in the `--classes-count parameter`. For imbalanced datasets with binary classification the weight multiplier can be set to 1 for class 0 and to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) for class 1. Format: ``` <value for class 1>,..,<values for class N> ``` For example: ``` 0.85,1.2,1 ``` Alert Do not use this parameter with auto\_class\_weights. Type - list - dict - collections.OrderedDict Default value None (the weight for all classes is set to 1) Supported processing units CPU and GPU ## class\_names #### Description Classes names. Allows to redefine the default values when using the MultiClass and Logloss metrics. If the upper limit for the numeric class label is specified, the number of classes names should match this value. Warning The quantity of classes names must match the quantity of classes weights specified in the `--class-weights` parameter and the number of classes specified in the `--classes-count` parameter. Format: ``` <name for class 1>,..,<name for class N> ``` For example: ``` smartphone,touchphone,tablet ``` Type list of strings Default value None Supported processing units CPU and GPU ## auto\_class\_weights Command-line: `--auto-class-weights` #### Description Automatically calculate class weights based either on the total weight or the total number of objects in each class. The values are used as multipliers for the object weights. Supported values: - None — All class weights are set to 1 - Balanced: C W k \= m a x c \= 1 K ( ∑ t i \= c w i ) ∑ t i \= k w i CW\_k=\\displaystyle\\frac{max\_{c=1}^K(\\sum\_{t\_{i}=c}{w\_i})}{\\sum\_{t\_{i}=k}{w\_{i}}} - SqrtBalanced: C W k \= m a x c \= 1 K ( ∑ t i \= c w i ) ∑ t i \= k w i CW\_k=\\sqrt{\\displaystyle\\frac{max\_{c=1}^K(\\sum\_{t\_i=c}{w\_i})}{\\sum\_{t\_i=k}{w\_i}}} Alert Do not use this parameter with `class_weights` and `scale_pos_weight`. Type string Default value None — All class weights are set to 1 Supported processing units CPU and GPU ## scale\_pos\_weight #### Description The weight for class 1 in binary classification. The value is used as a multiplier for the weights of objects from class 1. Note For imbalanced datasets, the weight multiplier can be set to ( s u m \_ n e g a t i v e s u m \_ p o s i t i v e ) \\left(\\frac{sum\\\_negative}{sum\\\_positive}\\right) Alert Do not use this parameter with `auto_class_weights` and `class_weights`. Type float Default value 1\.0 Supported processing units CPU and GPU ## boosting\_type Command-line: `--boosting-type` #### Description Boosting scheme. Possible values: - Ordered — Usually provides better quality on small datasets, but it may be slower than the Plain scheme. - Plain — The classic gradient boosting scheme. Type string Default value Depends on the processing unit type, the number of objects in the training dataset and the selected learning mode - CPU Plain - GPU - Any number of objects, MultiClass or MultiClassOneVsAll mode: Plain - More than 50 thousand objects, any mode: Plain - Less than or equal to 50 thousand objects, any mode but MultiClass or MultiClassOneVsAll: Ordered Supported processing units CPU and GPU Only the Plain mode is supported for the MultiClass loss on GPU ## boost\_from\_average Command-line: `--boost-from-average` #### Description Initialize approximate values by best constant value for the specified loss function. Sets the value of bias to the initial best constant value. Available for the following loss functions: - RMSE - Logloss - CrossEntropy - Quantile - MAE - MAPE Type bool Default value Depends on the selected loss function: - True for RMSE, Quantile, MAE, MAPE - False for all other loss functions Supported processing units CPU and GPU ## langevin Command-line: `--langevin` #### Description Enables the Stochastic Gradient Langevin Boosting mode. Refer to the [SGLB: Stochastic Gradient Langevin Boosting](https://arxiv.org/abs/2001.07248) paper for details. Type bool Default value False Supported processing units CPU ## diffusion\_temperature Command-line: `--diffusion-temperature` #### Description The diffusion temperature of the Stochastic Gradient Langevin Boosting mode. Only non-negative values are supported. Type float Default value 10000 Supported processing units CPU ## posterior\_sampling Command-line: `--posterior-sampling` #### Description If this parameter is set several options are specified as follows and model parameters are checked to obtain uncertainty predictions with good theoretical properties. Specifies options: - `Langevin`: true, - `DiffusionTemperature`: objects in learn pool count, - `ModelShrinkRate`: 1 / (2. \* objects in learn pool count). Type bool Default value False Supported processing units CPU only ## allow\_const\_label Command-line: `--allow-const-label` #### Description Use it to train models with datasets that have equal label values for all objects. Type bool Default value False Supported processing units CPU and GPU ## score\_function Command-line: `--score-function` #### Description The [score type](https://catboost.ai/docs/en/concepts/algorithm-score-functions) used to select the next split during the tree construction. Possible values: - Cosine (do not use this score type with the Lossguide tree growing policy) - L2 - NewtonCosine (do not use this score type with the Lossguide tree growing policy) - NewtonL2 Type string Default value Cosine Supported processing units The supported score functions vary depending on the processing unit type: - GPU — All score types - CPU — Cosine, L2 ## monotone\_constraints Command-line: `--monotone-constraints` #### Description Impose monotonic constraints on numerical features. Possible values: - "1" — Increasing constraint on the feature. The algorithm forces the model to be a non-decreasing function of this features. - "\-1" — Decreasing constraint on the feature. The algorithm forces the model to be a non-increasing function of this features. - "0" — constraints are disabled. Supported formats for setting the value of this parameter (all feature indices are zero-based): - Set constraints individually for each feature as a string (the number of features is n). Format ``` "(<constraint_0>, <constraint_2>, .., <constraint_n-1>)" ``` Zero constraints for features at the end of the list may be dropped. In `monotone_constraints = "(1,0,-1)"`an increasing constraint is set on the first feature and a decreasing one on the third. Constraints are disabled for all other features. - Set constraints individually for each explicitly specified feature as a string (the number of features is n). ``` "<feature index or name>:<constraint>, .., <feature index or name>:<constraint>" ``` These examples ``` monotone-constraints = "2:1,4:-1" ``` ``` monotone-constraints = "Feature2:1,Feature4:-1" ``` are identical, given that the name of the feature index 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set constraints individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<constraint_0>, <constraint_2>, .., <constraint_n-1>] ``` ``` {"<feature index or name>":<constraint>, .., "<feature index or name>":<constraint>} ``` Array examples ``` monotone_constraints = [1, 0, -1] ``` These dictionary examples ``` monotone_constraints = {"Feature2":1,"Feature4":-1} ``` ``` monotone_constraints = {"2":1, "4":-1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list of strings - string - dict - list Default value Python package, R package None Command-line Ommited Supported processing units CPU ## feature\_weights Command-line: `--feature-weights` #### Description Per-feature multiplication weights used when choosing the best split. The score of each candidate is multiplied by the weights of features from the current split. Non-negative float values are supported for each weight. Supported formats for setting the value of this parameter: - Set the multiplication weight for each feature as a string (the number of features is n). Format ``` "(<feature-weight_0>,<feature-weight_2>,..,<feature-weight_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Multiplication weights equal to 1 at the end of the list may be dropped. In this example ``` feature_weights = "(0.1,1,3)" ``` the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the multiplication weight individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<weight>, .., <feature index or name>:<weight>" ``` Note Spaces between values are not allowed. These examples ``` feature_weights = "2:1.1,4:0.1" ``` ``` feature_weights = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the multiplication weight individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-weight_0>, <feature-weight_2>, .., <feature-weight_n-1>] ``` ``` {"<feature index or name>":<weight>, .., "<feature index or name>":<weight>} ``` Array examples ``` feature_weights = [0.1, 1, 3] ``` These dictionary examples ``` feature_weights = {"Feature2":1.1,"Feature4":0.3} ``` ``` feature_weights = {"2":1.1, "4":0.3} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 1 for all features Supported processing units CPU ## first\_feature\_use\_penalties Command-line: `--first-feature-use-penalties` #### Description Per-feature penalties for the first occurrence of the feature in the model. The given value is subtracted from the score if the current candidate is the first one to include the feature in the model. Refer to the [Per-object and per-feature penalties](https://catboost.ai/docs/en/concepts/algorithm-score-functions) section for details on applying different score penalties. Non-negative float values are supported for each penalty. - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] ``` ``` {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} ``` Array examples. `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = [0.1, 1, 3] ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = [0.1, 1, 3] ``` These dictionary examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` first_feature_use_penalties = {"2":1.1, "4":0.1} ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` per_object_feature_penalties = {"2":1.1, "4":0.1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 0 for all features Supported processing units CPU ## fixed\_binary\_splits Command-line: `--fixed-binary-splits` #### Description A list of indices of binary features to put at the top of each tree; ignored if `grow_policy` is `Symmetric`. Type list Default value None Supported processing units GPU ## penalties\_coefficient Command-line: `--penalties-coefficient` #### Description A single-value common coefficient to multiply all penalties. Non-negative values are supported. Type float Default value 1 Supported processing units CPU ## per\_object\_feature\_penalties Command-line: `--per-object-feature-penalties` #### Description Per-object penalties for the first use of the feature for the object. The given value is multiplied by the number of objects that are divided by the current split and use the feature for the first time. Refer to the [Per-object and per-feature penalties](https://catboost.ai/docs/en/concepts/algorithm-score-functions) section for details on applying different score penalties. Non-negative float values are supported for each penalty. Python package - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". - Set the penalty individually for each required feature as an array or a dictionary (the number of features is n). Format ``` [<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>] ``` ``` {"<feature index or name>":<penalty>, .., "<feature index or name>":<penalty>} ``` Array examples. `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = [0.1, 1, 3] ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = [0.1, 1, 3] ``` These dictionary examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` first_feature_use_penalties = {"2":1.1, "4":0.1} ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = {"Feature2":1.1,"Feature4":0.1} ``` ``` per_object_feature_penalties = {"2":1.1, "4":0.1} ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". R package - Set the penalty for each feature as a string (the number of features is n). Format ``` "(<feature-penalty_0>, <feature-penalty_2>, .., <feature-penalty_n-1>)" ``` Note Spaces between values are not allowed. Values should be passed as a parenthesized string of comma-separated values. Penalties equal to 0 at the end of the list may be dropped. Penalties equal to 0 at the end of the list may be dropped. In this example `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "(0.1,1,3)" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "(0.1,1,3)" ``` Note Spaces between values are not allowed. the multiplication weight is set to 0.1, 1 and 3 for the first, second and third features respectively. The multiplication weight for all other features is set to 1. - Set the penalty individually for each explicitly specified feature as a string (the number of features is n). Format ``` "<feature index or name>:<penalty>,..,<feature index or name>:<penalty>" ``` Note Spaces between values are not allowed. These examples `first_feature_use_penalties` parameter: ``` first_feature_use_penalties = "2:1.1,4:0.1" ``` ``` first_feature_use_penalties = "Feature2:1.1,Feature4:0.1" ``` `per_object_feature_penalties` parameter: ``` per_object_feature_penalties = "2:1.1,4:0.1" ``` ``` per_object_feature_penalties = "Feature2:1.1,Feature4:0.1" ``` are identical, given that the name of the feature indexed 2 is "Feature2" and the name of the feature indexed 4 is "Feature4". Type - list - numpy.ndarray - string - dict Default value 0 for all objects Supported processing units CPU ## model\_shrink\_rate Command-line: `--model-shrink-rate` #### Description The constant used to calculate the coefficient for multiplying the model on each iteration. The actual model shrinkage coefficient calculated at each iteration depends on the value of the `--model-shrink-mode`for the Command-line version parameter. The resulting value of the coefficient should be always in the range (0, 1\]. Type float Default value The default value depends on the values of the following parameters: - `--model-shrink-mode` for the Command-line version - `--monotone-constraints` for the Command-line version Supported processing units CPU ## model\_shrink\_mode Command-line: `model_shrink_mode` #### Description Determines how the actual model shrinkage coefficient is calculated at each iteration. Possible values: - Constant: 1 − m o d e l \_ s h r i n k \_ r a t e ⋅ l e a r n i n g \_ r a t e , 1 - model\\\_shrink\\\_rate \\cdot learning\\\_rate {,} - m o d e l \_ s h r i n k \_ r a t e model\\\_shrink\\\_rate is the value of the `--model-shrink-rate`for the Command-line version parameter. - l e a r n i n g \_ r a t e learning\\\_rate is the value of the `--learning-rate`for the Command-line version parameter - Decreasing: 1 − m o d e l \_ s h r i n k \_ r a t e i , 1 - \\frac{model\\\_shrink\\\_rate}{i} {,} - m o d e l \_ s h r i n k \_ r a t e model\\\_shrink\\\_rate is the value of the `--model-shrink-rate`for the Command-line version parameter. - i i is the identifier of the iteration. Type string Default value Constant Supported processing units CPU
Shard	169 (laksa)
Root Hash	17435841955170310369
Unparsed URL	ai,catboost!/docs/en/references/training-parameters/common s443