Query about the SFT Data and Code used in Section4.5 Training-free vs. SFT

[rookie-zgy1513]

Thank you sincerely for your great work.！
As mentioned in Section 4.5 Training-free vs. SFT of your paper in Table 5, using a small subset of TimeChat-Online-139K and VideoChatOnline-IT datasets for training has achieved significant performance improvements on OVO-Bench. I am very interested in the data used for this SFT and the specific training script. Could you please make the data used for SFT and the specific training code public? I would be extremely grateful!

[YiwengXie]

Hi, @rookie-zgy1513

Thank you for your interest!

As mentioned in the paper, we used a subset of TimeChat-Online-139K and VideoChatOnline-IT for SFT with a standard training pipeline. The improvement on OVO-Bench mainly comes from the HLD metric, which benefits from negative samples during training that encourage the model to say it cannot answer when visual evidence is insufficient.

[rookie-zgy1513]

Thank you sincerely for your reply! This also answers my questions about the HLD metric during the reproduction process. Thank you very much!🙏