Apppublic

pufanyi/vbvrpro_output

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

App README

VBVR-Pro Output Explorer

Interactive task- and sample-level visualization of the strict In-Domain DanceGRPO sweep for Wan2.2 TI2V-5B.

—5 checkpoints × 4 sampling modes
—100 EvalKit-supported tasks and 500 scored samples per run
—SFT epoch-1 baseline comparison at both task and sample level
—Every scored 1024×1024, 161-frame video, paired with its exact main_v2 score

The static frontend and score indexes live in this Space. The 3.67 GiB media archive is hosted in the companion public dataset `pufanyi/vbvrpro_output-data` and streamed through the Hub CDN.

The strict-sweep results use split-manifest SHA-256 326f7bda3743e9c66dc0c29445661a5dda4ad0cee4cb8838c3fcfd0c4a149deb and VBVR EvalKit main_v2 revision 42a1593d8e493370c768be8e43646f0e0a9d8525.

“Out-of-Domain” follows the original VBVR benchmark naming. Those task families were present in this RL training dataset, so the label must not be interpreted as zero-shot task generalization. The 500 scored benchmark samples are sample-disjoint from the manifest training IDs.

This Space is generated from the authoritative score JSON files by scripts/eval/vbvr_pro/build_vbvrpro_space.py in Wan-Trainer.