G. Zang

icon indicating a website link https://www.hit.edu.cn/

icon company Harbin Institute of Technology icon location China icon location A student from HIT, focusing on recommendation systems and multimodal algorithms.

Results 3 issues of G. Zang

[Bug] internvl的MPO的loss一开始就为0

1 comment

### Checklist - [x] 1. I have searched related issues but cannot get the expected help. - [ ] 2. The bug has not been fixed in the latest version....

Can I add the module using your llava-next code?

Thank you for your work! I would like to add a multi-class classification layer after the LLM. Is it possible to directly add it using your code?

Why format reward is always zero?

Thanks for you excellent work! I train RL for 20 steps using your cold-start model, but the fornat reward is always 0.