EchoMinkki/Orbdivination
Thyme Video RL — pass-rate filtered subset RL training data for a video agent that first skims 64 uniformly sampled frames, then calls a temporal-crop tool to pull higher-FPS segments from the source mp4. Each row therefore carries both: videos — the 64 overview frames, embedded as JPEG bytes video_path — the source mp4, which the crop tool reads at rollout time Provenance Built from 12,000 multiple-choice questions over 1,479 videos (LVBench, LongVideoBench… See the full description on the dataset page: https://huggingface.co/datasets/EchoMinkki/Orbdivination.
Update videos_manifest.tsv: videos are referenced, not distributed
Update README.md: videos are referenced, not distributed
Remove build summary: contained internal absolute paths
Add pass-rate filtered video RL subset (5384 rows)
initial commit