Hi, thank you for the great work and for open-sourcing the dataset!
I noticed that the dataset README (mid_training_video/) explicitly lists the following files as part of the dataset composition:
mid_training_video/caption_v0/split_180s.jsonl — Captions for 180-second video clips
mid_training_video/caption_v0/split_gt10min.jsonl — Captions for >10-minute video clips
The viewer configs also reference preview parquets for these subsets (viewer_caption_180s, viewer_caption_gt10min).
However, when downloading from the HuggingFace repo (mvp-lab/LLaVA-OneVision-2-Data), only the following caption files are available:
✅ caption_v0/split_30s.jsonl (4.2M entries)
✅ caption_v0/split_60s.jsonl (2.7M entries)
❌ caption_v0/split_180s.jsonl — not found
❌ caption_v0/split_gt10min.jsonl — not found
Similarly, the mapping files (mapping_180s.csv with 714K entries and mapping_gt10min.csv with 347K entries) are present, but the corresponding video WebDataset tar shards (180s/ and gt10min/ directories) are missing.
The paper mentions that the mid-training video data includes clips of 30–60s, 60–180s, and 10–15min durations. Could you please clarify:
Are split_180s.jsonl and split_gt10min.jsonl planned for release?
Will the corresponding video tar shards (for 180s and >10min clips) also be uploaded?
If not planned, is there a recommended way to reproduce these captions from the mapping [CSVs?]
Thank you!
Hi, thank you for the great work and for open-sourcing the dataset!
I noticed that the dataset README (mid_training_video/) explicitly lists the following files as part of the dataset composition:
mid_training_video/caption_v0/split_180s.jsonl — Captions for 180-second video clips
mid_training_video/caption_v0/split_gt10min.jsonl — Captions for >10-minute video clips
The viewer configs also reference preview parquets for these subsets (viewer_caption_180s, viewer_caption_gt10min).
However, when downloading from the HuggingFace repo (mvp-lab/LLaVA-OneVision-2-Data), only the following caption files are available:
✅ caption_v0/split_30s.jsonl (4.2M entries)
✅ caption_v0/split_60s.jsonl (2.7M entries)
❌ caption_v0/split_180s.jsonl — not found
❌ caption_v0/split_gt10min.jsonl — not found
Similarly, the mapping files (mapping_180s.csv with 714K entries and mapping_gt10min.csv with 347K entries) are present, but the corresponding video WebDataset tar shards (180s/ and gt10min/ directories) are missing.
The paper mentions that the mid-training video data includes clips of 30–60s, 60–180s, and 10–15min durations. Could you please clarify:
Are split_180s.jsonl and split_gt10min.jsonl planned for release?
Will the corresponding video tar shards (for 180s and >10min clips) also be uploaded?
If not planned, is there a recommended way to reproduce these captions from the mapping [CSVs?]
Thank you!