ChatGLM-6B icon indicating copy to clipboard operation
ChatGLM-6B copied to clipboard

[Help] 训练数据必须一问一答吗,能不能喂跑团记录那种一长串的数据?

Open TEENN2233 opened this issue 2 years ago • 2 comments

Is there an existing issue for this?

  • [X] I have searched the existing issues

Current Behavior

网上的数据集都是一问一答的,让大模型像跑团DM一样围绕着一个主题,不断讨论,并让事情逐渐发展,实测这个能力有,但是该如何通过微调提升相关场景下的水平呢

Expected Behavior

No response

Steps To Reproduce

如题

Environment

- OS:windows11
- Python: 3.8
- Transformers:4.26.1
- PyTorch:1.12
- CUDA Support:True

Anything else?

No response

TEENN2233 avatar Apr 19 '23 01:04 TEENN2233

自己弄成多轮对话。

cywjava avatar Apr 19 '23 08:04 cywjava

对于Corpus[n] 的内容的生成, context 是 {Corpus_[i] | i<n} 吗? 那直接把context := {Corpus_[i] | i<n} 作为 prompt整个塞进去好了 如果我问题理解的正确的话

AlexLiu-C avatar Apr 20 '23 03:04 AlexLiu-C

https://github.com/THUDM/ChatGLM-6B/tree/main/ptuning#%E5%AF%B9%E8%AF%9D%E6%95%B0%E6%8D%AE%E9%9B%86

duzx16 avatar Apr 21 '23 12:04 duzx16