AnchorBenchmark/ANCHORDataset

Inside the release · Get started · Annotations · Video access · Citation

Joel Julin* · Souraja Kundu* · Liza Dahiya · George Z Wei · Liuyue Xie ·
Jingru Yi · Xu Zhang · Hao Yang · Laszlo A. Jeni
_{* Equal contribution · Carnegie Mellon University & Amazon AGI}

01 / Inside this release

Does a model really see and hear what it claims? ANCHOR connects audio-visual question answering to the evidence and reasoning behind an answer. The full benchmark combines questions, temporal evidence, spatial grounding, and machine-assisted, human-refined reasoning.

This repository provides the annotation and video-retrieval export. Original videos are retrieved from their upstream sources rather than redistributed here.

2,319 Video records

8,272 Question–answer entries

A benchmark built around evidence

[t] When?	[x, y] Where?	[♪] What can be heard?
Temporal intervals identify relevant moments in this export.	The full benchmark links objects and regions to reasoning; spatial annotations are not included in this export.	Audio-visual reasoning traces describe relevant speech and acoustic events where available.

Explore the interactive taxonomy and annotated video examples to see the full benchmark in context.

Dataset composition · exact counts in this export

Source label	Video records	QA entries
AVQA	751	751
Animal	163	798
COIN	86	86
Ego4D	106	130
LVBench	12	168
Maverix	568	1,835
MovieChat	491	2,455
THUMOS14	107	1,941
VideoMME	25	75
Unknown	10	33
Total	2,319	8,272

Animal, Maverix, THUMOS14, and VideoMME are the labels used in the download metadata. The paper names these sources Animal-Bench, MAVERIX, THUMOS-14, and Video-MME. Ten records retain an Unknown source label; their provenance is not inferred here.

The paper organizes questions into Information Query, Sports, Gaming, Shopping, Sentiment, Egocentric Agent, and Social Interaction. Per-question taxonomy labels are not included in this export.

02 / Get started

Download the annotations

python -m pip install -U huggingface_hub
hf download AnchorBenchmark/ANCHORDataset --repo-type dataset --local-dir ANCHOR

Or download only the annotation and retrieval files in Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AnchorBenchmark/ANCHORDataset",
    repo_type="dataset",
    local_dir="ANCHOR",
    allow_patterns=["data.jsonl", "video_metadata.json", "download_videos.py"],
)

Read your first question

Each JSONL row represents one video, with its questions nested under actionAnnotationList.

import json
from pathlib import Path

root = Path("ANCHOR")
metadata = json.loads((root / "video_metadata.json").read_text())

with (root / "data.jsonl").open() as stream:
    video_record = json.loads(next(stream))

annotation = video_record["annotation"]
video_path = annotation["video"]["src"]
qa = annotation["actionAnnotationList"][0]

print("Source:", metadata[video_path]["dataset"])
print("Video:", video_path)
print("Interval:", qa["start"], qa["end"])
print("Question:", qa["question"])
print("Answer:", qa["answer"])
reasoning = qa["gemini_analysis"]
if isinstance(reasoning, dict):
    reasoning = reasoning.get("text", "")
print("Reasoning:", reasoning)

The release preserves the original annotation structure, including mixed metadata types and edited reasoning objects. Reading with Python's json module avoids assuming a uniform tabular schema.

03 / Understand the annotations

File	Contents
data.jsonl	One video record per line; nested questions, answers, intervals, and available reasoning traces.
video_metadata.json	Retrieval metadata keyed by the exact annotation.video.src path.
download_videos.py	Upstream video retrieval helper; supports source filtering and dry runs.
release_stats.json	Counts computed directly from this export.
checksums.json	SHA-256 checksums for the three original payload files.

Annotation field reference

Path	Meaning
version	Annotation format version.
annotation.video.src	Relative video path, also the key into video_metadata.json.
annotation.video.fps, frames, duration, height, width	Exported video metadata. Values may be missing, zero, or placeholders; probe retrieved videos before relying on them.
annotation.actionAnnotationList	The video's list of QA annotations.
…question / …answer	Question and reference answer.
…start / …end	Exported segment boundaries in seconds. Check against the retrieved video's timeline before slicing.
…description	Original QA description, which may include multiple-choice options.
…gemini_analysis	Exported reasoning trace: a string or a structured object with text, editing metadata, and optional version history. Text may be empty.
annotation.keyframeList	Present but empty in this export.
annotation.objectAnnotationListMap	Present but empty in this export; no bounding-box tracks are distributed here.
annotation.regionAnnotationListMap / skeletonAnnotationListMap	Present but empty in this export.

QA entries also retain action, object, and color fields from the annotation format. Their presence does not imply that corresponding spatial tracks are included.

Video-retrieval metadata

Each value in video_metadata.json includes:

• dataset: source-dataset label.
• video_id: exported video identifier.
• download_source: retrieval route, such as youtube, huggingface, or ego4d.
• download_id: upstream ID or file reference.
• start_sec / end_sec: optional source-clip boundaries.
• extra: source-specific retrieval information.

Join this metadata to annotations using the full video path, rather than assuming that basenames are unique.

04 / Access the videos

Install the retrieval dependencies and make ffmpeg available on your system, then run the helper from the downloaded directory:

cd ANCHOR
python -m pip install yt-dlp datasets huggingface_hub tqdm

# Inspect the download plan before retrieving media.
python download_videos.py --dry_run

# Start with a small source subset.
python download_videos.py --datasets LVBench COIN VideoMME --output_dir ./videos

Route	What to expect
YouTube	Retrieved with yt-dlp; availability depends on the original upload. Some clips require trimming with ffmpeg.
Hugging Face	Source repositories may require separate access approval and authentication.
Ego4D	Obtain access through Ego4D, download the original files, and pass --ego4d_dir /path/to/ego4d_videos.
Animal-Bench	Videos are retrieved from upstream archives; the helper may need large cached archives.
Manual / unknown	Some items need manual retrieval or corrected source metadata. The helper cannot guarantee automatic access to every record.

The helper skips existing files and supports --workers 8 for increased parallelism. Its source-specific requirements and manual placement instructions are documented in the script. Upstream media may become unavailable independently of this release.

Evaluate a model

Use the ANCHOR evaluation repository for model inference, multi-judge answer and reasoning evaluation, token-based conciseness, and judge-agreement analysis. Resolve video paths locally and check reference traces before using them as evaluation targets.

Intended use and limitations

ANCHOR supports research on audio-visual question answering and evidence-grounded reasoning. Read answer quality alongside evidence support and reasoning alignment; a plausible explanation alone is not evidence that a model used the correct cues.

This is an export of the annotation and retrieval files, not a complete redistribution of all paper artifacts. Missing traces, absent spatial annotations, incomplete video metadata, and unknown source labels should be accounted for in any experiment. Machine-assisted traces may contain errors or fail to match a question; they should not all be treated as human-verified ground truth. No train/validation/test assignment is declared in these files.

License and upstream rights

The ANCHOR annotation release is licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0). Attribution is required; the license permits noncommercial sharing of the licensed material and does not permit sharing adapted material. See the full license for its terms.

Original video content remains subject to each upstream dataset's licenses and access conditions. Access to these annotations does not grant additional rights to third-party media. Please cite the relevant source datasets alongside ANCHOR.

Citation

@inproceedings{julin2026anchor,
  title={ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark},
  author={Julin, Joel and Kundu, Souraja and Dahiya, Liza and Wei, George Z and
          Xie, Liuyue and Yi, Jingru and Zhang, Xu and Yang, Hao and Jeni, Laszlo A.},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}

Keep reasoning connected to evidence.
Explore ANCHOR → · Questions & discussion