[BUG/Help] bias in self identification in role playing tasks

[fangzhangmnm]

Is there an existing issue for this?

• [X] I have searched the existing issues

Current Behavior

The model keep identifying itself as a 猫娘 no matter what character it is supposed to play after I finetuned it for roleplaying tasks using the ptuning script in this repo There is no cat in the training data. I think it might be a bias in the base model

Expected Behavior

The model identify itself as the character It learned from the training data.

Steps To Reproduce

Prepare training data. I use chatgpt to convert a novel to a drama script, like {prompt="A: dialogue B: " label="dialogue"} ~30K training items. Use the ptuning script in the repo to train the model. Let it complete the text "A: 请介绍一下你自己 B:"

Environment

- OS:
- Python:
- Transformers: 
- PyTorch:
- CUDA Support (`python -c "import torch; print(torch.cuda.is_available())"`) :

Anything else?

No response