在chat上SFT max_seq_len开大后指令遵循能力下降很严重

menghonghan commented 8 months ago

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

[X] 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

[X] 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

请问在chat上全量SFT max_seq_len开到32k相比2048 指令遵循能力下降非常严重请问可能有什么原因呢是数据不够长？谢谢解答谢谢

期望行为 | Expected Behavior

No response

复现方法 | Steps To Reproduce

No response

运行环境 | Environment

- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

No response

WangJianQ-cmd commented 7 months ago

您说的这个指令遵从能力，可以举一个具体的例子吗？

jklj077 commented 6 months ago

Directy training on long sequences can be harmful to model performances, as in training, dynamic_ntk and logn attention is effectively disabled; see line 833 of modeling_qwen.py. It means that the model can only handle a sequence length up to 2K or 8K at the start of the finetuning and adapting to a 32K sequence length will require a lot of data.

We advise you to take a look at Qwen1.5, which changes the strategy in extending context length and is more admissive to finetuning on long sequences.

QwenLM / Qwen