Skip to content

Roaoch/CyberClassic-trainer

Repository files navigation

CyberClassic-trainer

This is a training environment for model of CyberClassic collection. Current training environment pipline contain three steps: FineTune GPT2 model, FineTune T5 model, Reinforcement learning of text generator Env has two separete datasets

  • True dataset. Size 13048 rows. Column: Text - single sentence from the texts of Dostovesky F.M.
  • False dataset. Size 5771 rows. Column: Text - single sentence from the texts of Kuprin A.I. and sentences geenerated with RuGPT3

FineTune GPT2 model

On this step base GPT2 model finetuned on true dataset.

FineTune T5 model

On this step we choose from true dataset 6000 rows, add new colunt "labels" with values 1 and 0 to true dataset and false dataset respectively, then contcatenet them. In the end we have model for binary classification of text sequence by belonging to style of Dostovesky F.M.

Reinforcement learning

On this step we perform second round of training text generation model, with TRL dependencie. Reward function is a simple socer from classifier multiplied by 10.

Part of CyberClassic model

Trainer enviroment for ML-modle of telegram bot

HuggingFace Collection