rugpt3medium_based_on_gpt2

Model was trained with sequence length 1024 using transformers lib by SberDevices 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.

Authors

• NLP core team RnD Telegram channel:
  • Dmitry Zmitrovich