Benwei Lu

似乎您提供的格式并不是sharegpt格式的数据集,和官方提供的有些出入,这是自定义的数据集格式么?LLaMA-Factory能够支持么? It seems that the format you offered is not the offical sharegpt format, how can i use through LLaMA-Factory? Do i need to modify the `dataset_info.json` in LLaMA-Factory?

Thanks for your jobs, we are looking forward for your work.

SFT时似乎没有给出明确的微调格式,还支持定位框送到模型里微调吗?

> > SFT时似乎没有给出明确的微调格式,还支持定位框送到模型里微调吗? > > 试了几种方式，数据量在3k，效果都很差 那看来是不支持bbox微调了,官方到现在也没说

似乎您提供的格式并不是sharegpt格式的数据集,和官方提供的有些出入,这是自定义的数据集格式么?LLaMA-Factory能够支持么? It seems that the format you offered is not the offical sharegpt format, how can i use through LLaMA-Factory? Do i need to modify the dataset_info.json in LLaMA-Factory?/cc @kq-chen

请问你是怎么微调的呢?用的LLaMA-Factory吗?他这个给的格式好像不是标准的sharegpt格式,你是怎么实现的呀?@Guangming92

> Thank you for your support of Qwen2_VL. The main reason for the poor fine-tuning grounding performance is that the 1D rope position embedding, rather than MRoPE, was used during...

Have you solved? I check the numpy version and it is correct.

> Have you solved? I check the numpy version and it is correct. I followed [zth15804074521](https://github.com/zth15804074521)'s advice, uninstall then reinstall with `pip install numpy==1.26.4` , and it works!