Request for the release of caption files for datasets generated by GPT-4.1 and video subtitle generation code.

[likanchuan09171]

Hi, thank you for your excellent work and for sharing your research with the community!

I’m very interested in reproducing the results from your paper. Would it be possible to release the caption files produced for each dataset as described in the paper?

Regenerating the caption database from scratch can be quite expensive, and due to the stochastic nature of LLM generation, it may also result in differences that affect reproducibility. Having access to the original caption files would be extremely helpful for accurate benchmarking and further research.

Thank you very much for considering this request!

[anonymous626]

hi @likanchuan09171 we've uploaded the database files including the captions and subject registries. Let me know what else I can help.

[likanchuan09171]

Thank you so much for your quick response and for sharing the database files! This is extremely helpful for my research and for the community. I really appreciate your support and openness.

[anhdao69]

@anonymous626 Did you use the full set of videos from EgoSchema, LongVideoBench, and VideoMME for evaluation? In the caption JSON file, I only found 753 videos for LongVideoBench, even though the benchmark has over 3,000. I noticed a similar issue with EgoSchema and VideoMME, while LVBench looks fine.

[likanchuan09171]

Hi @anonymous626, thanks again for sharing the caption data!

Could you also release the code or scripts used to transcribe video into .srt subtitles?
It helps ensure reproducibility and consistency.
Thanks a lot for considering this!

Best

[anonymous626]

hi @Johnx69 sorry for the late reply. We are using the validation set for both EgoSchema and LongVideoBench. For LongVideoBench valset, there are 753 videos wherein 188 of them are in long set. For EgoSchema val set, there are 500 videos. For VideoMME, we only evaluate on the long set as paper suggests.

Thanks for pointing out and we will emphasize correspondingly in our recent updated paper.

[anonymous626]

hi @likanchuan09171 sorry for the late reply. I've update here. It is very close to the original example of whisper x.