More details in fine-tuning with MotionVid-QA

[Junhojuno]

Hello, authors!

I have some questions about fine-tuning the VLM following the paper.
The sentences "During SFT, the vision tower, LLM, and merger components were all trainable. During DPO training, we exclusively trained the LLM part." in Appendix D are not clear to me. I guess LLM is LoRA-tuned, but vision tower is not sure. is the vision tower also LoRA tuned? or fully fine-tuned?

Additionally, it is also not clear the fine-tuning setting, such as epochs.

Could you please help clarify these?

[natsunoshion]

Sorry for the late response and thank you for your interest in our work!

We used full-parameter fine-tuning (except for the frozen parts) with 3 epochs. For more specific configurations we used, we have uploaded the ft_config to the repository for reference.