Zhang Chen

> install tensorflow 1.14 solved my problem, thank you

Seems davinci haven't generated sentence with `"` as a EOS, when I add `--keep_outputs_without_eos` flag, everything is OK.

I met the same error, and solved just by setting `OPENAI_PROXY` in config.yaml. Sees the http_proxy/https_proxy in env not work.

> This is a good idea. Would you mind shooting us an email at [[email protected]](mailto:[email protected]). I'd like to chat a bit over email about your use case so we can...

> which finetune.py were you using? did you mix the use of Qwen and Qwen2 model and code? I'm using [Qwen/finetune.py](https://github.com/QwenLM/Qwen/blob/main/finetune.py). I was finetuning Qwen 1.5 with model_max_length 32768, and...

> we are also in the process of deprecating the finetune.py in this repo and we advise you to use training frameworks, including [Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl), [Llama-Factory](https://github.com/hiyouga/LLaMA-Factory), [Swift](https://github.com/modelscope/swift), etc., to finetune your...

@yangjianxin1 where can I found `src/train.py` mentioned in https://qwen.readthedocs.io/en/latest/training/SFT/llama_factory.html? 

how should I set the param `--flash_attn`  There is an error msg "**train.py: error: argument --flash_attn: expected one argument**"

according to [README](https://github.com/QwenLM/Qwen/blob/main/README.md), Qwen-7B does support 32K Max Length. 

I noticed input_ids are padded to the max_length in batch, but labels are not. The checkpoint is at `site-packages\transformers\tokenization_utils_base.py`, line 3450   ``` Traceback (most recent call last): File...