Classification models in DeepPavlov

In DeepPavlov one can find code for training and using classification models which are implemented as a number of different neural networks or sklearn models. Models can be used for binary, multi-class or multi-label classification. List of available classifiers (more info see below):

  • BERT classifier (see here) builds BERT 4 architecture for classification problem on PyTorch.

  • PyTorch classifier (see here) builds neural network on PyTorch.

  • Sklearn classifier (see here) builds most of sklearn classifiers.

Quick start

Command line

INSTALL First whatever model you have chose you would need to install additional requirements:

python -m deeppavlov install <path_to_config>

where <path_to_config> is a path to one of the provided config files or its name without an extension, for example “insults_kaggle_bert”.

To download pre-trained models, vocabs, embeddings on the dataset of interest one should run the following command providing corresponding name of the config file (see above) or provide flag -d for commands like interact, train, evaluate:

python -m deeppavlov download  <path_to_config>

where <path_to_config> is a path to one of the provided config files or its name without an extension, for example “insults_kaggle_bert”.

INTERACT One can run the following command to interact in command line interface with provided config:

python -m deeppavlov interact <path_to_config> [-d]

where <path_to_config> is a path to one of the provided config files or its name without an extension, for example “insults_kaggle_bert”. With the optional -d parameter all the data required to run selected pipeline will be downloaded.

TRAIN After preparing the config file (including change of dataset, pipeline elements or parameters) one can train model from scratch or from pre-trained model optionally. To train model from scratch one should set load_path to an empty or non-existing directory, and save_path to a directory where trained model will be saved. To train model from saved one should set load_path to existing directory containing model’s files (pay attention that model can be loaded from saved only if the clue sizes of network layers coincide, other parameters of model as well as training parameters, embedder, tokenizer, preprocessor and postprocessors could be changed but be attentive in case of changing embedder - different embeddings of tokens will not give the same results). Then training can be run in the following way:

python -m deeppavlov train <path_to_config>

where <path_to_config> is a path to one of the provided config files or its name without an extension, for example “insults_kaggle_bert”. With the optional -d parameter all the data required to run selected pipeline will be downloaded.

Python code

One can also use these configs in python code.

INTERACT To download required data one have to set download parameter to True. Then one can build and interact a model from configuration file:

from deeppavlov import build_model

model = build_model('insults_kaggle_bert', download=True)  # in case of necessity to download some data

model = build_model('insults_kaggle_bert', download=False)  # otherwise

print(model(["You are dumb", "He lay flat on the brown, pine-needled floor of the forest"]))

>>> ['Insult', 'Not Insult']

TRAIN Also training can be run in the following way:

from deeppavlov import train_model

model = train_model('insults_kaggle_bert', download=True)  # in case of necessity to download some data

model = train_model('insults_kaggle_bert', download=False)  # otherwise

BERT models

BERT (Bidirectional Encoder Representations from Transformers) 4 is a Transformer pre-trained on masked language model and next sentence prediction tasks. This approach showed state-of-the-art results on a wide range of NLP tasks in English.

deeppavlov.models.torch_bert.torch_transformers_classifier.TorchTransformersClassifierModel (see here) provides easy to use solution for classification problem using pre-trained BERT. Several pre-trained English, multi-lingual and Russian BERT models are provided in our BERT documentation.

Two main components of BERT classifier pipeline in DeepPavlov are deeppavlov.models.preprocessors.torch_transformers_preprocessor.TorchTransformersPreprocessor and deeppavlov.models.torch_bert.torch_transformers_classifier.TorchTransformersClassifierModel (see here). The deeppavlov.models.torch_bert.torch_transformers_classifier.TorchTransformersClassifierModel class supports any Transformer-based model.

Non-processed texts should be given to torch_transformers_preprocessor for tokenization on subtokens, encoding subtokens with their indices and creating tokens and segment masks. If one processed classes to one-hot labels in pipeline, one_hot_labels should be set to true.

torch_transformers_classifier has a dense layer of number of classes size upon pooled outputs of Transformer encoder, it is followed by softmax activation (sigmoid if multilabel parameter is set to true in config).

Neural Networks on PyTorch

deeppavlov.models.classifiers.TorchClassificationModel (see here) could be used for implementation of different neural network configurations for classification task.

If you want to build your own architecture for text classification tasks, do the following:

from deeppavlov.models.classifiers.torch_classification_model import TorchTextClassificationModel

class MyModel(TorchTextClassificationModel):

    def my_network_architecture(self, **kwargs):
        model = <create Torch model using parameters from kwargs>
        return model

In the config file, assign "class_name": "" and "model_name": "my_network_architecture" in the dictionary with the main model.

If you want to build your own PyTorch-based model for some other NLP task, do the following:

from deeppavlov.core.models.torch_model import TorchModel

class MyModel(TorchModel):

    def train_on_batch(x, y, *args, **kwargs):
        <your code here>
        return loss

    def __call__(data, *args, **kwargs):
        <your code here>
        return predictions

    def my_network_architecture(self, **kwargs):
        model = <create Torch model using parameters from kwargs>
        return model

In the config file, assign "class_name": "" and "model_name": "my_network_architecture" in the dictionary with the main model.

Sklearn models

deeppavlov.models.sklearn.SklearnComponent (see here) is a universal wrapper for all sklearn model that could be fitted. One can set model_class parameter to full name of model (for example, sklearn.feature_extraction.text:TfidfVectorizer or sklearn.linear_model:LogisticRegression). Parameter infer_method should be set to class method for prediction (predict, predict_proba, predict_log_proba or transform). As for text classification in DeepPavlov we assign list of labels for each sample, it is required to ensure that output of a classifier-sklearn_component is a list of labels for each sample. Therefore, for sklearn component classifier one should set ensure_list_output to true.

Pre-trained models

We also provide with pre-trained models for classification on “AG News” dataset, “Detecting Insults in Social Commentary”, Twitter sentiment in Russian dataset.

Detecting Insults in Social Commentary dataset contains binary classification task for detecting insults for participants of conversation. Train, valid and test division is the same as for the Kaggle challenge.

AG News dataset contains topic classification task for 5 classes (range from 0 to 4 points scale). Test set is initial one from a web-site, valid is a Stratified division 1/5 from the train set from web-site with 42 seed, and the train set is the rest.

Twitter mokoron dataset contains sentiment classification of Russian tweets for positive and negative replies 1. It was automatically labeled. Train, valid and test division is made by hands (Stratified division: 1/5 from all dataset for test set with 42 seed, then 1/5 from the rest for validation set with 42 seed). Two provided pre-trained models were trained on the same dataset but with and without preprocessing. The main difference between scores is caused by the fact that some symbols (deleted during preprocessing) were used for automatic labelling. Therefore, it can be considered that model trained on preprocessed data is based on semantics while model trained on unprocessed data is based on punctuation and syntax.

RuSentiment dataset contains sentiment classification of social media posts for Russian language within 5 classes ‘positive’, ‘negative’, ‘neutral’, ‘speech’, ‘skip’.

SentiRuEval dataset contains sentiment classification of reviews for Russian language within 4 classes ‘positive’, ‘negative’, ‘neutral’, ‘both’. Datasets on four different themes ‘Banks’, ‘Telecom’, ‘Restaurants’, ‘Cars’ are combined to one big dataset.

Questions on Yahoo Answers labeled as either informational or conversational dataset contains intent classification of English questions into two category: informational (0) and conversational (1) questions. The dataset includes some additional metadata but for the presented pre-trained model only Title of questions and Label were used. Embeddings were obtained from language model (ELMo) fine-tuned on the dataset

L6 - Yahoo! Answers Comprehensive Questions and Answers. We do not provide datasets, both are available upon request to Yahoo Research. Therefore, this model is available only for interaction.

Stanford Sentiment Treebank contains 5-classes fine-grained sentiment classification of sentences. Each sentence were initially labelled with floating point value from 0 to 1. For fine-grained classification the floating point labels are converted to integer labels according to the intervals [0, 0.2], (0.2, 0.4], (0.4, 0.6], (0.6, 0.8], (0.8, 1.0] corresponding to very negative, negative, neutral, positive, very positive classes.









Insult detection



English BERT




1.1 Gb



5-classes SST on conversational BERT




1.1 Gb


Twitter mokoron


RuWiki+Lenta emb w/o preprocessing




6.2 Gb


Multilingual BERT




1.3 Gb

Conversational RuBERT



1.5 Gb


DeepPavlov Topics


Distil BERT base uncased

F1-w / F1-m



0.7 Gb

GLUE Benchmark

The General Language Understanding Evaluation (GLUE) benchmark is a collection of resources for training, evaluating, and analyzing natural language understanding systems. More details are on the official page

In DeepPavlov there is a set of configuration files to run training and evaluation on GLUE tasks train/dev sets. DeepPavlov (DP) results on dev sets are averaged over 3 runs. We report the same metrics as on the official leaderboard










DP bert-base-cased









DP bert-base-uncased









HuggingFace bert-base-uncased









How to train on other datasets

We provide dataset reader BasicClassificationDatasetReader and dataset BasicClassificationDatasetIterator to work with .csv and .json files. These classes are described in readers docs and dataset iterators docs.

Data files should be in the following format (classes can be separated by custom symbol given in the config as class_sep, here class_sep=","):











To train model one should

  • set data_path to the directory to which train.csv should be downloaded,

  • set save_path to the directory where the trained model should be saved,

  • set all other parameters of model as well as embedder, tokenizer and preprocessor to desired ones.

Then training process can be run in the same way:

python -m deeppavlov train <path_to_config>

How to improve the performance

  • One can use FastText 2 to train embeddings that are better suited for considered datasets.

  • One can use some custom preprocessing to clean texts.

  • One can use ELMo 3 or BERT 4.

  • All the parameters should be tuned on the validation set.



Ю. В. Рубцова. Построение корпуса текстов для настройки тонового классификатора // Программные продукты и системы, 2015, №1(109), –С.72-78

  1. Bojanowski, E. Grave, A. Joulin, T. Mikolov, Enriching Word Vectors with Subword Information.


Peters, Matthew E., et al. “Deep contextualized word representations.” arXiv preprint arXiv:1802.05365 (2018).


Devlin J. et al. Bert: Pre-training of deep bidirectional transformers for language understanding //arXiv preprint arXiv:1810.04805. – 2018.