VideoPRM V6 — current accepted snapshot

This private research snapshot contains 27229 unique accepted QA, comprising 1,600 unchanged V5 SFT warm-up records and 25629 RL records. The original V6 target is 30,000; 2771 remain uncollected. This is an audited current snapshot, not a completed 30,000-example release. The independent export audit withheld 17 historical accepted Object records exceeding the original 14-per-physical-visit cap; all source records and media remain preserved locally. See audit/withheld_from_export.json. Data collection may continue independently.

V5 warm-up checkpoint compatibility

The parent is VideoVPR/VideoPRM-CanonicalV5-Data at revision 1d4c7f4539e94c0569e27c42c2ac41cdf2b6e3b4. data/sft_warmup.jsonl is copied byte-for-byte: SHA256 e1359e2a6a65b87f065ab6cbe865b7d921f02b52b8a83110fef2b09a0f2a023e. All 1,600 IDs, order, fields, question/answer content and frame-sampling settings remain identical. Every referenced warm-up video payload is unchanged; the V5 media archives are reused byte-for-byte. Warm-up membership is never resampled or changed to 800. The V5 data files are exact prefixes of the corresponding V6 files.

Reuse the existing SFT checkpoint trained on this V5 warm-up data for the subsequent RL stage. Do not include warm-up rows again in the RL policy input. No training was launched by this export, and checkpoint weights are not included or modified.

Task counts

Task	Accepted	Original target	Remaining
relative_direction	1947	4200	2253
egocentric_temporal_direction	4200	4200	0
route_plan	4700	4700	0
object_grounded_route_planning	2109	2400	291
route_shape	4700	4700	0
reverse_route_plan	4624	4700	76
event_order	2149	2300	151
before_or_after_action	2800	2800	0

Files and usage

• data/all.jsonl: full supervised master; includes answers and provenance, for audit/SFT use.
• data/sft_warmup.jsonl: the exact V5 1,600-record warm-up file.
• data/grpo_policy.jsonl: RL policy inputs only, without hidden targets or assistant answers.
• data/grpo_targets.jsonl: aligned hidden verifier targets, separate from policy inputs.
• data/grpo_structured_gold.jsonl: aligned full RL training records with structured gold.
• media/videos-*.tar: unchanged V5 archives; media/v6-added-videos-*.tar: additional V6 videos.
• media/video_index.jsonl: SHA-addressed media index; exact V5 index prefix retained.
• catalog/train.parquet: browsing index, with training_stage; not the answer-free RL policy view.
• audit/warmup_v5_identity.json: exact warm-up and media compatibility proof.
• audit/validation.json, audit/task_remaining.json: automated audit and honest shortfalls.
• audit/visual_review80/: 80 real fixed24 contact sheets; human decisions are blank.

from huggingface_hub import snapshot_download
root = snapshot_download("YilingMa/V6", repo_type="dataset")
print(root)

python /path/to/snapshot/tools/prepare_data.py --root /path/to/snapshot --output ./videoprm_v6

The preparation tool verifies checksums and media and resolves portable videos/ references into runtime paths. Relative warm-up file bytes are identical to V5; absolute runtime prefixes depend on the chosen extraction directory.

Audit and access

Original schema/reward-one, source geometry/visibility/fixed24, benchmark physical/visit/content isolation, source Training splits, global QA/new-semantic dedup, physical-scene caps and policy/hidden-target separation remain enforced. Inherited frozen V5 records remain unchanged, including any reported legacy-only semantic exceptions. The audit does not claim per-example human review. Source raw data, failures and partial outputs remain preserved locally and are not added as unaccepted QA.

This private dataset contains derived media from EPIC-KITCHENS-100, ScanNet, ARKitScenes and R2R/RxR/Matterport3D. Applicable source-dataset access and redistribution terms remain in effect. No new unapproved source has been included.