Natural video only; no media changes.
- Corrected 138 answers.
- Removed 5 ambiguous SoccerNet questions.
- Removed 85 repeated questions (same question and label on byte-identical footage as a kept copy).
- Added 4 questions: "Did the person clean a counter or surface in this video?" (answer yes, L=512): mq_B4_v18_q1, mq_B4_v28_q3, mq_B4_v80_q3, mq_B4_v88_q3.
- Added question_mode "exists" to every question in manifest.json.
- Rows: 1,813 -> 1,727 (designed 1,028 -> 1,007, density_topup 785 -> 720). Manifest video entries: 1,028 -> 1,015 (13 entries left with no questions are dropped; their mp4s are unchanged).
- README.md and metadata.json counts, sizes and checksums updated to match.
Merge natural_video_density into natural_video. The two configs described the same 1028 clips with disjoint question sets, so they are now one config of 1813 rows, with a `question_set` column ('designed' / 'density_topup') telling them apart; natural_video_density is deleted. No media moves. The merge also brings the designed questions up to the same notation: source_provenance is now on all 1813 rows (950 EPIC rows shipped null), question_time is length_L rather than 0.0, and density_band, event timing, clip_sha256, query_class and label_source are populated wherever the source supports it. Fixes clip_sha256, which was resolved by clip basename: 637 basenames are reused across the per-length directories, so 465 rows carried the hash of a different length's file, breaking the documented group-on-clip_sha256 recipe. README, LICENSES and Croissant metadata regenerated for the single natural config (2.2.0).
Fix a batch of dataset defects found in an adversarial review. natural_video_density: manifest.json now carries questions so main.py can load it; add media_config so video_path resolves (this config reuses natural_video's clips); add integer event_start/end_frame and event_timing (measured for EPIC, centered_by_construction for SoccerNet); add clip_sha256 since 89 groups of clips are byte-identical under different names; drop constant margin_s/question_mode; lowercase source_dataset to match natural_video and the licence texts. natural_video: rewrite SoccerNet provenance, whose event position was clip-relative and whose negative windows were placeholders, so the documented reconstruction recipe now works. LICENSES: synthetic-and-easyhuman.txt named files that no longer exist; SoccerNet-NOTE.txt rewritten and corrected. Top-level license is now 'other' (mixed) rather than cc-by-4.0 over NC/NDA content. Croissant: correct descriptions, add FileSets for the mp4s, bump to 2.1.0. Adds media_config to ALL configs so one documented join rule resolves media everywhere.
Fix a batch of dataset defects found in an adversarial review. natural_video_density: manifest.json now carries questions so main.py can load it; add media_config so video_path resolves (this config reuses natural_video's clips); add integer event_start/end_frame and event_timing (measured for EPIC, centered_by_construction for SoccerNet); add clip_sha256 since 89 groups of clips are byte-identical under different names; drop constant margin_s/question_mode; lowercase source_dataset to match natural_video and the licence texts. natural_video: rewrite SoccerNet provenance, whose event position was clip-relative and whose negative windows were placeholders, so the documented reconstruction recipe now works. LICENSES: synthetic-and-easyhuman.txt named files that no longer exist; SoccerNet-NOTE.txt rewritten and corrected. Top-level license is now 'other' (mixed) rather than cc-by-4.0 over NC/NDA content. Croissant: correct descriptions, add FileSets for the mp4s, bump to 2.1.0.
natural_video_density: switch density_band to codec distortion (785 rows, L64-L1024). The previous labels were consec_cos_dist, now superseded, and disagreed with the codec band on 205 of 880 rows; L=16 was dropped with the protocol change. Updates README and Croissant metadata to describe the codec protocol.
natural_video_density: switch density_band to codec distortion (785 rows, L64-L1024). The previous labels were consec_cos_dist, now superseded, and disagreed with the codec band on 205 of 880 rows; L=16 was dropped with the protocol change. Updates README and Croissant metadata to describe the codec protocol.
Add nested manifest.json per config for direct main.py ingestion (co-exists with parquet+NDJSON); fix synthetic_video input_sequence by joining with text-config S_tokens
