climba/VideoGen-LLM-Wan-REPA-adapters

VideoGen-LLM Wan-REPA Adapter Checkpoints

Selected VideoChat2-HD adapter checkpoints from Wan-REPA video understanding experiments.

These are not full base models. They are adapter checkpoints produced by experiments/videochat2_hd_wan_repa_finetune.py and require the corresponding VideoChat2-HD/Mistral setup used in the repository.

Included Checkpoints

Directory	Setting	Result
eq_seed127_lambda0p1/	Equivariance Wan-REPA, seed 127, lambda 0.1	49/96 on 5-fold MotionBench subset
eq_seed123_lambda0p1/	Equivariance Wan-REPA, seed 123, lambda 0.1	46/96 on 5-fold MotionBench subset
relation_only/	Wan temporal relation alignment, seed 123	46/96 on 5-fold MotionBench subset

Notes

• The full local workspace is on GitHub: https://github.com/gustn9609/VideoGen-LLM
• Base model weights, HDF5 feature caches, and raw videos are not included here.
• The later negative-control and larger-set experiments showed that the Wan-specific gain is weak in the current setup. See final_summary.md for the full interpretation.

Files

• adapter_checkpoint.pt: fold-wise adapter/repa-head checkpoint
• finetune_eval.md: evaluation table for that checkpoint
• sweep_summary.md: seed/lambda robustness summary
• final_summary.md: final experiment conclusion