Fixed best audio-visual model

OCS-DETR fixed best model

Text-guided video moment retrieval and highlight detection, built on QD-DETR and further developing TR-DETR's HD–MR interaction idea.

The fixed audio-visual checkpoint reproduces all 14 saved summary metrics over 1,550 QVHighlights validation queries: MR average mAP 43.63; R1@0.5 64.13; R1@0.7 48.06.

Assets contain tensor-only weights and an eight-query feature sample. Run python scripts/download_assets.py, then python scripts/evaluate.py --smoke --device cpu. Sizes and SHA-256 hashes are recorded in the repository's asset manifest.