|
--- |
|
license: mit |
|
language: |
|
- ru |
|
metrics: |
|
- perplexity |
|
- bleu |
|
- rouge |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
This text generator is based on OpenAI GPT2 model from HuggingFace |
|
Base model went through two step of learning |
|
## First - Finetining of base model |
|
On this step model is finetuned on dataset of single sentence from the texts of Dostovesky F.M. |
|
|
|
Training parameters: |
|
* Epoch = 10 |
|
* Learning Rate = 1e-3 |
|
* Optimizer = AdamW |
|
* Scheduler = OneCycleLR |
|
* Training env = PyTorch |
|
|
|
![image.png](https://cdn-uploads.huggingface.co/production/uploads/645ced0dce72244df7b0f6d2/Mvp2ouBpVkK8t4CfuBvx7.png) |
|
|
|
![image.png](https://cdn-uploads.huggingface.co/production/uploads/645ced0dce72244df7b0f6d2/_AiPUPQOi5-raYSyZNohh.png) |
|
|
|
## Second - RL |
|
On this step finetuned model went trough reinforcement learning pipline with TRL library. |
|
|
|
Training parameters: |
|
* Epoch = 30 |
|
* Trainer = PPO |
|
* Query texts = first 100 texts from dataset, trimmed by first 3 words |
|
* Reward = score from [binary classifier](https://huggingface.co/Roaoch/CyberClassic-Discriminator) multiplied by 10 |
|
|
|
![image.png](https://cdn-uploads.huggingface.co/production/uploads/645ced0dce72244df7b0f6d2/Mv_8ZAfml6tPcdXBOXcQG.png) |
|
|
|
|
|
![image.png](https://cdn-uploads.huggingface.co/production/uploads/645ced0dce72244df7b0f6d2/ZbeI0Tq6eSy-_Tl_gK-sz.png) |