Home

Awesome

Russian GPT-3 models

ruGPT3XL, ruGPT3Large, ruGPT3Medium, ruGPT3Small and ruGPT2Large

This repository contains bunch of autoregressive transformer language models trained on a huge dataset of russian language.

Table of contents

ruGPT3XL

Setup

For colab we recommend use the following installation instructions:

export LD_LIBRARY_PATH=/usr/lib/
apt-get install clang-9 llvm-9 llvm-9-dev llvm-9-tools
git clone https://github.com/qywu/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
pip install triton
DS_BUILD_CPU_ADAM=1 DS_BUILD_SPARSE_ATTN=1 pip install deepspeed
pip install transformers
pip install huggingface_hub
pip install timm==0.3.2
git clone  https://github.com/sberbank-ai/ru-gpts
cp ru-gpts/src_utils/trainer_pt_utils.py /usr/local/lib/python3.8/dist-packages/transformers/trainer_pt_utils.py
cp ru-gpts/src_utils/_amp_state.py /usr/local/lib/python3.8/dist-packages/apex/amp/_amp_state.py

After installation env please restart colab. For checking is all ok, run the following commands:

!ds_report
# Output:
...
sparse_attn ............ [YES] ...... [OKAY]
...
import deepspeed.ops.sparse_attention.sparse_attn_op

Usage

Here is a simple example of usage. For more see this example or Open In Colab.

import sys
from src.xl_wrapper import RuGPT3XL
import os

# If run to from content root.
sys.path.append("ru-gpts/")
os.environ["USE_DEEPSPEED"] = "1"
# We can change address and port
os.environ["MASTER_ADDR"] = "127.0.0.1"
os.environ["MASTER_PORT"] = "5000"
gpt = RuGPT3XL.from_pretrained("sberbank-ai/rugpt3xl", seq_len=512)
gpt.generate(
    "Кто был президентом США в 2020? ",
    max_length=50,
    no_repeat_ngram_size=3,
    repetition_penalty=2.,
)

Finetuning

Example of finetune, load finetuned model and generate is here.

Our example of finetuning script here

Pretraining details ruGPT3XL

Model was trained with 512 sequence length using Deepspeed and Megatron code by Devices team, on 80B tokens dataset for 4 epochs. After that model was finetuned 1 epoch with sequence length 2048.
Note! Model has sparse attention blocks.

Total training time was around 10 days on 256 GPUs.
Final perplexity on test set is 12.05.

🤗HuggingFace model card link.

ruGPT3Large, ruGPT3Medium, ruGPT3Small, ruGPT2Large

Setup

For using ruGPT3Large, ruGPT3Medium, ruGPT3Small, ruGPT2Large just install 🤗HuggingFace transformers.

pip install transformers==4.24.0

Usage

Here we can obtain examples of finetuning or generation.

Also this examples is adapted for google colab:

from transformers import GPT2LMHeadModel, GPT2Tokenizer


model_name_or_path = "sberbank-ai/rugpt3large_based_on_gpt2"
tokenizer = GPT2Tokenizer.from_pretrained(model_name_or_path)
model = GPT2LMHeadModel.from_pretrained(model_name_or_path).cuda()
text = "Александр Сергеевич Пушкин родился в "
input_ids = tokenizer.encode(text, return_tensors="pt").cuda()
out = model.generate(input_ids.cuda())
generated_text = list(map(tokenizer.decode, out))[0]
print(generated_text)
# Output should be like this:
# Александр Сергеевич Пушкин родился в \n1799 году. Его отец был крепостным крестьянином, а мать – крепостной крестьянкой. Детство и юность Пушкина прошли в деревне Михайловское под Петербургом. В 1820-х годах семья переехала

Pretraining details

All pretraining was done on Nvidia Tesla V100-SXM3 32 Gb GPUs on a Christofari Cluster. Following are the details of pretraining for each model.

Pretraining details ruGPT3Large

Model was trained with sequence length 1024 using transformers lib by Devices team on 80B tokens for 3 epochs. After that model was finetuned 1 epoch with sequence length 2048.

Total training time was around 14 days on 128 GPUs for 1024 context and few days on 16 GPUs for 2048 context.
Final perplexity on test set is 13.6.

You can obtain this model by using transformers with model name sberbank-ai/rugpt3large_based_on_gpt2.

🤗HuggingFace model card link

Our pretraining script here

Pretraining details ruGPT3Medium

Model was trained with sequence length 1024 using transformers lib by Devices team on 80B tokens for 3 epoch. After that model was finetuned on 2048 context.

Total training time was around 16 days on 64 GPUs.
Final perplexity on test set is 17.4.

You can obtain this model by using transformers with model name sberbank-ai/rugpt3medium_based_on_gpt2.

🤗HuggingFace model card link

Our pretraining script here

Pretraining details ruGPT3Small

Model was trained with sequence length 1024 using transformers by Devices team on 80B tokens around 3 epoch. After that model was finetuned on 2048 context.

Total training time took around one week on 32 GPUs.

You can obtain this model by using transformers with model name sberbank-ai/rugpt3small_based_on_gpt2.

🤗HuggingFace model card link

Our pretraining script here

Pretraining details ruGPT2Large

Model was trained with sequence length 1024 using transformers by Devices team on 170Gb data on 64 GPUs 3 weeks.

You can obtain this model by using transformers with model name sberbank-ai/rugpt2large.

🤗HuggingFace model card link

OpenSource Solutions with ruGPT3

Papers mentioning ruGPT3

According to google scholar search - feel free to add links to this list

Text Simplification

@article{shatilovsentence,
  title={Sentence simplification with ruGPT3},
  author={Shatilov, AA and Rey, AI},
  url={http://www.dialog-21.ru/media/5281/shatilovaaplusreyai142.pdf}
}

@article{fenogenovatext,
  title={Text Simplification with Autoregressive Models},
  author={Fenogenova, Alena},
  url={http://www.dialog-21.ru/media/5250/fenogenovaa141.pdf}}

Text Detoxification

@article{dementieva2021methods,
  title={Methods for Detoxification of Texts for the Russian Language},
  author={Dementieva, Daryna and Moskovskiy, Daniil and Logacheva, Varvara and Dale, David and Kozlova, Olga and Semenov, Nikita and Panchenko, Alexander},
  journal={arXiv preprint arXiv:2105.09052},
  year={2021},
  url={https://arxiv.org/abs/2105.09052}
}

Paraphrasing and Data Augmentation

@inproceedings{fenogenova2021russian,
  title={Russian Paraphrasers: Paraphrase with Transformers},
  author={Fenogenova, Alena},
  booktitle={Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing},
  pages={11--19},
  year={2021},
  url={https://www.aclweb.org/anthology/2021.bsnlp-1.2.pdf}
}

Model Evaluation

@article{malykh2021morocco,
  title={MOROCCO: Model Resource Comparison Framework},
  author={Malykh, Valentin and Kukushkin, Alexander and Artemova, Ekaterina and Mikhailov, Vladislav and Tikhonova, Maria and Shavrina, Tatiana},
  journal={arXiv preprint arXiv:2104.14314},
  year={2021},
  url={https://arxiv.org/abs/2104.14314}}