RoVidX_filtered

# RoVidX_filtered A quality-filtered subset of **RoVid-X** (DAGroup-PKU) — **500,000** robot-manipulation video clips selected for video-generation training. ## Contents - `prompts.jsonl` — one line per clip: `{"video_path": "video/<id>.mp4", "prompt": <short_caption>}` - `video/shard_00000.tar … shard_00249.tar` — the 500k `.mp4` clips packed into 250 tar shards (~1.7 GB each). Each tar member is `video/<id>.mp4`. ## Materialize the videos Extract the shards from the repo root — they reproduce the `video/` paths used in `prompts.jsonl`: ```bash for t in video/shard_*.tar; do tar -xf "$t"; done # -> video/<id>.mp4 , matching prompts.jsonl ``` ## How it was filtered (from ~3.03M RoVid-X clips) 1. has caption 2. frame length ∈ [49, 209] @30fps (≈1.6–7.0s) 3. semantic dedup of `short_caption` vs an in-house embodied prompt set (cosine ≥ 0.90) 4. technical visual-quality gating (decode-ok, not frozen, well-exposed) + ranking to the best 500k. Clips are 30 fps, high-resolution (min side ≥ 688 px).