SyncReward-Data-80K

This dataset contains:

• human audio-visual synchronization ratings of 57,099 generated audio-video clips;
• the clips themselves, as mp4 files in zip archives;
• the Stage-2 training splits derived from the ratings;
• SyncReward-Bench, a 1,000-clip benchmark for sync metrics.

It is the data behind the SyncReward reward model.

Paper, code repository and citation: coming soon.

Repository layout

annotations/annotations_80k.jsonl   80,001 ratings (one row per annotator x clip)
splits/stage2_train.jsonl           54,463 clips, Stage-2 training targets
splits/stage2_val.jsonl              1,636 clips, Stage-2 validation targets
bench/syncreward_bench.jsonl         1,000 clips, SyncReward-Bench
videos/<model>_partNNN.zip          57,099 mp4 files in 19 zip archives (213 GB)
videos/index.jsonl                  video_path -> zip archive
scripts/rebuild_targets.py          rebuilds splits/ from annotations/
SHA256SUMS                          checksums of all files except README.md

Download

hf download Guan123/SyncReward-Data --repo-type dataset --local-dir SyncReward-Data
cd SyncReward-Data
sha256sum -c SHA256SUMS
for z in videos/*.zip; do unzip -q -n "$z"; done      # -> generated/<model>/*.mp4

Every video_path in the jsonl files is relative to the extraction directory. Pass that directory as --video_root to the SyncReward code. To fetch only the metadata, or only one generator, use for example --include 'annotations/*' --include 'splits/*' --include 'bench/*' or --include 'videos/kling_*'. videos/index.jsonl tells you which archive holds a given clip.

from datasets import load_dataset
ratings = load_dataset("Guan123/SyncReward-Data", "annotations", split="train")
stage2 = load_dataset("Guan123/SyncReward-Data", "stage2")        # train / validation
bench = load_dataset("Guan123/SyncReward-Data", "bench", split="test")

Videos

Generator (model)	Videos	Size	Archives
davinci	6,691	7.95 GB	1
kling	6,718	47.43 GB	4
ltx2.3	6,772	18.23 GB	2
mova	6,661	30.42 GB	3
nava	6,796	23.17 GB	2
ovi	6,733	7.85 GB	1
seedance	6,728	24.30 GB	2
veo3	10,000	53.87 GB	4
Total	57,099	213.2 GB	19

The archives are uncompressed (stored) zips; mp4 is already compressed. Every clip referenced by any jsonl file is included. The clips were generated from text prompts by Kling 3.0, Seedance 2.0, Veo 3, LTX-2.3, daVinci, MOVA, NAVA and Ovi. Each generator's own terms of use also apply to its outputs (see License).

Training data

SyncReward-Data-80K contains 80,001 human synchronization ratings from 5 annotators over 57,099 clips generated by 8 text-to-audio-video models (about 10k ratings per model).

Each rating first marks whether the audio is related to the visible scene at all. If it is, the annotator scores synchronization as 0 (out of sync), 1 (partially synchronized) or 2 (well synchronized). We keep unrelated audio in the data on purpose: it is a common failure mode of current generators, and a sync reward should penalize it rather than never see it.

Rating	Count	Share
Unrelated audio	10,065	12.6%
0 – out of sync	12,907	16.1%
1 – partial	37,935	47.4%
2 – well synchronized	19,094	23.9%

The Stage-2 target of a clip is the rater-bias-corrected mean of its related ratings. Clips that every rater marked unrelated (6,261 train and 172 validation clips, 11.5% of the training pool) get target 0, the same as a clear sync failure. The 1,931 clips with mixed votes use their related ratings only.

Splits are clip-disjoint by (prompt, generator): 54,463 training and 1,636 validation clips. The 1,000 SyncReward-Bench clips are held out from training.

Fields

annotations: one row per (annotator, clip)

Field	Meaning
uid	prompt id; the same prompt is rendered by several generators
model	generator, see Videos
video_path	clip path relative to the extraction directory
rater	anonymized annotator, rater_1 ... rater_5
related	1 if the audio is related to the visible scene, 0 if not
score	0 / 1 / 2 sync score; null when related = 0
split	the clip's split: train, val or test (held out)
in_bench	the clip is in SyncReward-Bench

• By split: train has 54,463 clips and 75,609 ratings, val has 1,636 clips and 2,604 ratings, and test has 1,000 held-out clips and 1,788 ratings.
• Ratings per clip: 1 (38,695 clips), 2 (14,379), 3 (3,578), 4 (421), 5 (26).
• in_bench: marks 1,880 ratings. These are the original in-pool ratings; the bench target comes from an independent re-annotation.

stage2: one row per clip

Field	Meaning
video_path, uid, model	as above
target	Stage-2 regression target in [0, 2]
related	0 if every rater marked the audio unrelated (then target = 0), else 1

bench: one row per clip

Field	Meaning
id	bench id, curated1k_0000 ... curated1k_0999
video_path, uid, model	as above
target	mean of up to 3 independent human scores (0/1/2)
n_labels	number of scores averaged: 3 for 933 clips, 2 for 54, 1 for 13
prompt	text prompt used to generate the clip

SyncReward-Bench covers 7 generators (no Veo 3), with 142 or 143 clips each.

Rebuilding the Stage-2 targets

python scripts/rebuild_targets.py --check recomputes splits/stage2_*.jsonl from annotations/annotations_80k.jsonl and confirms the result is byte-identical to the released files.

1. Rater bias b_j = mean of (rater score - clip mean score) over clips with at least 2 scores: rater_1 -0.042, rater_2 -0.071, rater_3 -0.231, rater_4 +0.124, rater_5 +0.223.
2. Clip target = mean over the clip's related ratings of clip(score - b_j, 0, 2), rounded to 4 decimals.
3. Clips with no related rating get target = 0 and related = 0.

Splits and overlap

• Disjointness: splits are disjoint by (uid, model). The same prompt rendered by a different generator can still appear in several splits.
• Held-out clips: the 1,000 test clips were held out from training.
• Bench sources: of the 1,000 SyncReward-Bench clips, 817 come from test and 183 from val. The validation split was never used for gradient updates, but it did select the released checkpoint, so those 183 bench clips influenced model selection.
• Unused test clips: the other 183 test clips appear only in annotations.
• Training overlap: no bench clip is in stage2/train.

License

The annotations, manifests and scripts are released under Apache-2.0 (see LICENSE). The videos were produced by the third-party generators listed above. By using them you also agree to follow each generator's terms of use. They are intended for research on audio-visual synchronization and video generation evaluation.

Citation

Coming soon.