Q-Prefer: prompt curation candidates

Current OpenVidHD RL conditions: screened v2 (2026-10-26)

Use OpenVidHD-RL-Screened-20261026-v2 for the screened training manifests.

Screening result	Original50K	Original25K subset
Accepted	42,343	21,654
Review only	5,348	2,781
Excluded	1,209	565
Fixed validation/test	1,100	0

The main train.json and paired JSONL files contain42,343 accepted conditions. original25k-train.json and its paired JSONL files contain21,654 accepted conditions from the previous25K subset. JSON sample lists, review/exclusion records and the complete report are included. Review/excluded rows are not training rows.

Both accepted pools cover 988 existing labeled clusters, rather than1000 balanced clusters;26 full-pool rows have null cluster labels. The screening isolates665 training candidates sharing a parsed parent-video identifier with fixed holdouts. Validation100/test1000 stay byte-identical; they share one parsed parent identifier with each other, as documented in the report.

Text/metadata screening only: this is not per-example visual prompt-image verification. Images are not bundled in this folder; original captions and pinned first-frame references are preserved. These are conditions for RL rollout or candidate generation, not generated videos or preference-labeled RM training data.

Viewer config openvidhd-i2v now uses the accepted full-pool manifests. The unscreened image source remains explicitly available as openvidhd-i2v-source-50k; its images and attribution are retained because the new manifests depend on them.

DiffusionART T2V prompts

Artifacts: DiffusionART-T2V-Prompts-curated. Run: clip-text-dfd-k1000-t095-20261019T051143Z.

50,000 source prompts -> 49,172 provisionally retained prompts and 828 near-duplicate removal candidates. Human review is pending. The CLIP text cosine threshold 0.95 has known false merges; removal candidates are not all confirmed duplicates. No quality ranking or final human selection is implied.

Source chain: VidProM -> DanceGRPO -> DiffusionART-Prompts at commit 9d947fb510d67a8e5334ec1d8c7d8927f7658c08. Prompt text and sample IDs are unchanged. Source manifest SHA256: b54fe7c10e7530d24e5f928f3f40f9d6c160191301631235463945575b499194.

This is a DFD-inspired text-feature adaptation, not a reproduction of its video-feature/human-selection curation. We use normalized CLIP ViT-B/32 text features, source-order greedy cosine filtering, and Euclidean K-means (1,000 clusters, 50 iterations, seed 42). The 0.95 greedy filtering threshold is an added adaptation, not a DFD paper rule. The analysis does not measure video visual or motion diversity.

The run includes candidate JSONL manifests, cluster assignments, cluster and pair review pages, report figures, compact similarity records and provenance. Feature matrices and centroids were removed after processing; see cleanup.json. The T2V subset does not include videos or reference images. No model weights are distributed.

OpenVidHD I2V prompts and first frames

Artifacts: OpenVidHD-I2V-Prompts-50k.

50,000 distinct source clips from a completed 94,795-video CPU-quality-filtered OpenVidHD pool (at least 81 frames). Captions are matched exactly by original filename to OpenVid-1M, revision d8a63bd22989c80b5734ec2bb989f4e1b61a5807; they are not rewritten. These examples do not reconstruct DiffusionART's original I2V split.

Sampling is without replacement with weight 1 / measured video duration, seed 42. Each selected clip contributes its actual first decoded display frame. Mean source duration is 7.3191 s, versus 8.9438 s in the eligible pool. Images retain original resolution, encoded as JPEG q=2 without crop or resize.

Like DiffusionART's I2V JSONL, each row has prompt and sample_id; this release also includes image (relative first-frame path), archive/member paths, hashes, dimensions, source duration and quality provenance. Stable OpenVidHD-derived IDs identify this new dataset. Extract the 100 frames/train-*.tar archives into the subset directory to create first_frames/*.jpg alongside train.jsonl.

All 50,000 images and all uploaded shards passed SHA256 verification. Details, duplicate statistics and duration histograms are in manifest.json. There is no validation/test split and no additional semantic clustering. A complete release has UPLOAD_COMPLETE.json; source provenance and upstream notices are in provenance.json and attribution/. The OpenVid source card declares CC BY 4.0; original attribution and underlying media terms continue to apply.

Attribution and usage

Original notices and upstream provenance are preserved under DiffusionART-T2V-Prompts-curated/attribution/. This mirror does not grant additional rights to source material; consult those notices and source datasets. Prompts are user-generated and are not guaranteed to be safe or error-free.

Previous balanced sampling release (superseded)

The previous unscreened OpenVidHD-RL-Balanced25K-k1000-v1 folder has been removed from the current branch and replaced by the screened release above. Its original25K selection and provenance remain accessible at immutable commit c5420fb. Historical pinned revisions remain reproducible.