AV-Dense-60k

AV-Dense-60k contains paired video, audio, raw text, and dense audio-video captions. The media is packaged as deterministic WebDataset TAR shards, with portable metadata and checksums for streaming or restoration.

Rights and license notice: this repository uses license: other. Public availability does not relicense the underlying media and does not grant rights that the publisher does not hold. Before downloading, redistributing, or using any sample, users must independently confirm the media's provenance, copyright and neighboring rights, platform terms, privacy/consent requirements, and suitability for their intended jurisdiction and use.

Verified dataset inventory

Split	Records	Video	Audio	Intended role
train	44,178	44,178	44,178	model training
validation	5,522	5,522	5,522	held-out validation
test	5,523	5,523	5,523	final evaluation only

The split IDs are unique and pairwise disjoint. Every manifest ID has exactly one MP4 and one WAV in processed_gt_data; no missing, orphan, or zero-byte files were found. AV60KDenseBench.csv contains the same 5,523 IDs as test.jsonl and must not be used for training. The additional 480p_grpo variant contains the 5,522 validation IDs.

The upstream metadata contained obsolete machine-specific media locations. Release staging reconstructs every video/audio field from the split and video_id/id, so all structured paths in this repository are canonical and relative to the restored dataset root, for example processed_gt_data/train/videos/<video_id>.mp4.

Repository contents

This data repository contains only the following release surfaces:

README.md
metadata/
  train.jsonl
  val.jsonl
  test.jsonl
  AV60KDenseBench.csv
  AV69kDense-vanilla-max4s.pt
  align_csv.py
data/
  processed/train/avdense-train-*.tar
  processed/val/avdense-val-*.tar
  processed/test/avdense-test-*.tar
  legacy_480p_grpo/avdense-val480p-*.tar
manifests/
  shards.json
  SHA256SUMS

Training, extraction, and manifest-building code is distributed separately from this dataset repository. No model source, checkpoint, experiment report, or presentation is claimed as part of this public data release.

WebDataset sample format

Each processed sample has three same-stem members:

<video_id>.mp4
<video_id>.wav
<video_id>.json

The JSON member contains sample_id, split, variant, prompt_av, prompt_v, prompt_a, video_path, and audio_path. prompt_av/prompt_v use dense_caption; prompt_a uses raw_text; both media paths are relative.

Stream shards without extraction

Install the optional webdataset package, then run this example from the downloaded data repository root:

python -m pip install webdataset
python - <<'PY'
import json
from pathlib import Path

import webdataset as wds

shards = sorted(Path("data/processed/train").glob("avdense-train-*.tar"))
if not shards:
    raise SystemExit("no train shards found; run from the data repository root")

samples = wds.WebDataset([str(path) for path in shards], shardshuffle=False).to_tuple(
    "__key__", "mp4", "wav", "json"
)
for sample_id, mp4_bytes, wav_bytes, metadata_bytes in samples:
    metadata = json.loads(metadata_bytes)
    print(sample_id, len(mp4_bytes), len(wav_bytes), metadata["prompt_av"])
    break
PY

Integrity verification

Before extraction, run the checksum gate from the data repository root:

sha256sum --check manifests/SHA256SUMS

Do not extract or train if any shard is missing or is not reported as OK.

For a stronger, read-only audit, use extract_avdense_webdataset.py from the separately distributed JavisR1 code package. The --verify-only mode rechecks shard digests, manifest totals, TAR structure, sample counts, relative JSON paths, and every MP4/WAV/JSON triplet without writing output.

One reproducible relative workspace arrangement is:

workspace/
  AV-Dense-60k/          # this data repository
  code/jarvisr1/         # separately distributed code package
  datasets/              # restoration destination

From workspace/code/jarvisr1:

python3 scripts/extract_avdense_webdataset.py \
  --repo-root ../../AV-Dense-60k \
  --verify-only

Restore the training layout

Run restoration only after both checksum and verify-only gates succeed. From the same separately distributed code-package directory:

python3 scripts/extract_avdense_webdataset.py \
  --repo-root ../../AV-Dense-60k \
  --output-root ../../datasets/AV-Dense-60k

python3 scripts/build_avdense_manifests.py \
  --dataset-root ../../datasets/AV-Dense-60k

The extractor restores the normalized metadata plus the following default media tree:

../../datasets/AV-Dense-60k/
  train.jsonl
  val.jsonl
  test.jsonl
  processed_gt_data/train/{videos,audios}/...
  processed_gt_data/val/{videos,audios}/...
  processed_gt_data/test/{videos,audios}/...

To additionally restore the legacy validation-resolution variant, add --include-legacy-480p to the extraction command. The canonical builder produces:

sample_id,prompt_av,prompt_v,prompt_a,video_path,audio_path,split

Responsible use and limitations

• license: other is intentional; do not infer a permissive media license from public repository visibility.
• Users are responsible for rights clearance, attribution obligations, privacy review, consent, and compliance with applicable law and source-platform terms.
• Keep train, validation, and test IDs disjoint. Test and benchmark rows are evaluation only.
• Dense captions may contain factual, temporal, or synchronization errors and should not be treated as ground truth without task-appropriate review.
• Dataset inclusion does not imply endorsement of a depicted person, action, statement, or downstream application.
• This repository makes no model-quality claim.