rugpt3small_based_on_gpt2

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