InternVideo2 distilled CLIP models give poor retrieval performance — are fine-tuned text encoder weights available?

Hello,

I've been testing the InternVideo2 distilled CLIP models for zero-shot text-to-video retrieval and found that clip/S14, clip/B14, and clip/L14 from https://huggingface.co/OpenGVLab/InternVideo2_distillation_models/tree/main/clip give surprisingly poor results.

I'm aware that the distillation checkpoints are deltas that need to be applied on top of the
corresponding stage2 backbone weights — I'm loading them that way. The issue seems to be on the text encoder side: the instructions pair
them with mobileclip_blt.pt, but I suspect this may be the original base weights rather than the text encoder fine-tuned during CLIP
contrastive training.

For clip/L14, I noticed that https://huggingface.co/OpenGVLab/InternVideo2_CLIP_S is self-contained and includes text encoder weights,
which suggests a fine-tuned text encoder does exist for at least the L14 variant.

Could you confirm whether fine-tuned text encoder checkpoints exist for S14 and B14, and if so, whether they could be released?

Thanks for the great work on InternVideo2!

Start with the InternVideo2 distillation model instructions and compare the referenced Hugging Face checkpoints, including mobileclip_blt.pt and InternVideo2_CLIP_S. The issue is resolved when the maintainers confirm whether fine-tuned S14 and B14 text-encoder checkpoints exist and, if possible, publish them.