按照官方提供示例去微调chatglm6b之后，丧失了原有的通用能力，是什么原因，如何调整避免呢

wang-chunxing commented 11 months ago

Is there an existing issue for this?

[X] I have searched the existing issues

Current Behavior

[2023-11-09 10:17:50] ", prompt:"类型#上衣材质#牛仔布颜色#白色风格#简约图案#刺绣衣样式#外套衣款式#破洞", response:"'一款简约的牛仔外套，采用经典的牛仔面料，穿着舒适透气，搭配白色刺绣，时尚大气。而前襟的破洞设计，打破单调感，更具时尚感。'" INFO: 127.0.0.1:48516 - "POST / HTTP/1.1" 200 OK [2023-11-09 10:18:38] ", prompt:"你好", response:"'这件裙子是款，颜色很活泼，很年轻。整体是偏甜的风格，很显少女气息，搭配小衫，很显年轻。'" INFO: 127.0.0.1:49796 - "POST / HTTP/1.1" 200 OK

[2023-11-09 10:21:04] ", prompt:"你好是谁发明的", response:"'好享通的，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享通好享通，好享

Expected Behavior

除了回答特定领域的问题，还需要它的通用能力

Steps To Reproduce

按照p-tuning文件下的readme执行相关的[P-Tuning v2] 执行train.py

Environment

- OS: Linux
- Python: 3.9
- Transformers:
- PyTorch: 
- CUDA Support (`python -c "import torch; print(torch.cuda.is_available())"`) :

Anything else?

No response

jayus2 commented 7 months ago

+1

liubing0427 commented 4 months ago

+1

Roronoayx commented 3 months ago

请问你解决这个问题了吗？

THUDM / ChatGLM-6B