[Bug] Missing videofeatures for cmdad_eval

[AnnikaIgl]

Dear ant-research Team,

thank you for open-sourcing this code!
I’m trying to replicate your CMDAD results. I created the virtual environment and downloaded the checkpoints as described in the README, then ran the CMDAD evaluation command from the README.

During evaluation, the dataloader fails to load video features for 1318 out of 7316 samples.

After investigating, it looks like there is a mismatch between the CMDAD evaluation JSON and the provided video feature file:

• The downloaded file VideoLLaMa_CMD_eval_fp16.h5 (from your checkpoint on Hugging Face) contains features for 6769 AD entries.

• The CMDAD eval JSON contains 7316 AD entries.

Would it be possible to update the h5 file to include the features for all ADs of CMDAD or alternatively, share the steps/command(s) to generate the missing video features locally?

Thank you for your time!