YANG XINYU

question about degeneration problem

5

I reimplemented dpo in my own dataset and model, and the model tends to generate answers of increasing length:  Finally model generates long answers with repeated tokens. The same...