RVG-Bench / README.md
fyrisk159's picture
Align release prompt wording across modality and time variants
d2f9126 verified
|
Raw History Blame Contribute Delete
11.4 kB
metadata
pretty_name: RVG-Bench
license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - video
  - audio
  - temporal-grounding
  - reasoning
configs:
  - config_name: default
    data_files:
      - split: test
        path: annotations.jsonl

RVG-Bench

A Benchmark for Reasoning-Driven Audio-Visual Temporal Grounding

This package contains 1,890 question-answer annotations over 1,255 videos. Given a question and a video, the task is to identify the temporal interval that directly answers the question. This is an evaluation dataset; the test split contains all 1,890 annotations.

Project page · Paper: Coming soon · Evaluation code link: Coming soon

Annotation license: CC BY-NC-SA 4.0. Third-party videos and audio retain their original licenses and applicable terms; see Sources & license.

Access: The dataset is publicly available at Hugging Face.

Contents

.
├── README.md
├── LICENSE
├── LICENSES.md
├── annotations.jsonl
├── prompts/
└── videos/
    ├── Fun-QA/
    ├── daily/
    ├── videoholmes/
    └── worldsense/

annotations.jsonl contains one QA per line. Videos are stored once and can be referenced by several questions. Only videos referenced by the annotations are included. The MP4 files occupy 9,393,878,428 bytes (9.39 GB / 8.75 GiB) and preserve the original evaluation media bytes. All 1,255 files were checked to contain both video and audio streams. Video paths are relative to videos/.

Source directory Videos QA pairs
Fun-QA 142 187
daily 273 321
videoholmes 418 857
worldsense 422 525
Total 1,255 1,890

The QA modality labels are only visual (857), only audio (215), and both (818). There are 1,588 single-step questions and 302 multi-step questions. Multi-step questions contain 98 two-step, 203 three-step, and one four-step annotations.

Example: cups-and-ball magic

Query: What is the key deceptive action that secretly causes the ball to be under the cup for the final reveal?

Reference answer: The magician turns the cup upside down and places it on the table, but intentionally does not pour the ball out, trapping it inside.

Reference interval: 9–12 seconds. Category: Causal Reasoning. The corresponding video is videos/Fun-QA/M_C_23_0000_0755.mp4.

{
  "sample_id": "Fun-QA/M_C_23_0000_0755.json#0",
  "video_id": "Fun-QA/M_C_23_0000_0755",
  "video": "Fun-QA/M_C_23_0000_0755.mp4",
  "source": "Fun-QA",
  "question": "What is the key deceptive action that secretly causes the ball to be under the cup for the final reveal?",
  "reference_answer": "The magician turns the cup upside down and places it on the table, but intentionally does not pour the ball out, trapping it inside.",
  "start_sec": 9.0,
  "end_sec": 12.0,
  "category": "Causal Reasoning",
  "required_modality": "only visual",
  "duration": 27
}

Evaluation prompts

Prompt templates and usage guide provide Default CoT, Direct, Two-step, and Oracle Grounding templates for audio-visual and visual-only models, in absolute-seconds and MM:SS formats. A standalone Python helper renders the task text; model inference and media preparation use your own model interface. The templates use standardized release wording based on the experiment prompts; the guide documents their historical sources and adaptations.

The files in prompts/ use Apache-2.0; the annotation and third-party media licenses remain unchanged.

Download

The dataset repository is fyrisk159/RVG-Bench. The files can be downloaded without signing in.

pip install -U huggingface_hub

hf download fyrisk159/RVG-Bench \
  --repo-type dataset \
  --local-dir data/rvg-bench

To download only the annotations:

hf download fyrisk159/RVG-Bench annotations.jsonl \
  --repo-type dataset \
  --local-dir data/rvg-bench

Record the dataset commit hash for reproducible experiments. Use --revision COMMIT_OR_TAG to download a specific published revision. Private or gated repositories require access and authentication via hf auth login.

Annotation schema

Field Description
sample_id QA identifier fixed for this release: source-relative original JSON path followed by # and a zero-based QA index. These IDs match the evaluation loader applied to the final annotation collection.
video_id Source directory and video filename without .mp4.
video MP4 path relative to the videos/ directory.
source Source directory label.
question English question used as the model input in this dataset revision.
reference_answer Original reference text answer. Input only for the oracle Ground ablation.
start_sec, end_sec Reference answer interval, in absolute seconds from the start of the supplied video.
category Human-readable reasoning category. Source numeric prefixes are removed; k_hop_2, k_hop_3, and k_hop_4 are grouped as Multi-step Reasoning.
required_modality Original label: only visual, only audio, or both.
duration Original annotated video duration in seconds. This is not a replacement with a newly measured container duration.

The numeric time fields are converted from the original MM:SS annotations without clipping, rounding, or shifting the target interval. sample_id values identify questions rather than videos; keep them unchanged when saving predictions. Original per-video JSON files are not required to resolve an ID: use the video field to locate the media.

Keep the dataset revision with each experiment. Within this release, sample_id remains stable across wording refinements. Historical experiment files may use different within-video QA indices; verify their mapping against the annotations used for that run before assigning release IDs.

Read the annotations and resolve video paths using the Python standard library:

import json
from pathlib import Path

root = Path("data/rvg-bench")
with (root / "annotations.jsonl").open(encoding="utf-8") as handle:
    for line in handle:
        sample = json.loads(line)
        video_path = root / "videos" / sample["video"]
        question = sample["question"]
        # Supply video_path and question to your own model interface.

The Hugging Face configuration selects the JSONL annotations as the table for the test split. The video column is a relative path string, not an automatically decoded video feature. Download the MP4 files for local inference. Audio is accessed through the MP4 media rather than separate audio sidecar files.

Evaluation

Use the RVG-Bench code repository's prompt renderer and standalone scorer. Standard CoT, Direct, and Two-step evaluation use the question and the appropriate media input. Ground additionally receives reference_answer and must be reported as an oracle-answer ablation. Reference timestamps must not be supplied to the model.

Save one prediction per QA as JSONL, for example using the same ID as its annotation:

prediction = {
    "sample_id": sample["sample_id"],
    "start_sec": predicted_start_seconds,
    "end_sec": predicted_end_seconds,
}

From the evaluation code repository, with this dataset downloaded to data/rvg-bench:

python evaluation/eval_your_results.py \
  --annotations data/rvg-bench/annotations.jsonl \
  --predictions results/predictions.jsonl \
  --output results/scores.json

The scorer reports mIoU and IoU@0.3/0.5/0.7 as percentages. The default denominator is all 1,890 QA IDs; missing or malformed predictions receive zero IoU. This package supports category and modality breakdowns. Multi-step queries are identified by category: "Multi-step Reasoning"; per-query reasoning-step counts are not included. Inference is supplied by the user and is not part of this data package.

Annotation preservation notes

The release contains all 1,890 entries from the final annotation collection. It does not include the data construction pipeline, auxiliary model responses, or audit histories.

Question clarification (2026-10-16). For Fun-QA/M_C_23_0000_0755.json#0, “key action” was clarified to “key deceptive action” to match the project-page example. Its sample ID, video, reference answer, category, modality, and 9–12 second target interval are unchanged. Use the annotation revision associated with a run when reproducing its results.

One entry, videoholmes/jCQnUuq-TEE.json#1, has an inconsistent upstream is_valid=false flag while being included in the final collection. It is retained in this export; that flag is not used as an additional selection rule. Its status should be reconciled in a subsequent annotation review rather than silently changing the evaluation population.

Five entries have an annotated end time one second beyond their recorded duration:

Sample ID End (seconds) Recorded duration (seconds) Probed container duration (seconds)
videoholmes/qKTSekyZWh0.json#1 281 280 280.497042
worldsense/ExlUpRsr.json#0 39 38 38.139
worldsense/bLnoaJeS.json#0 99 98 98.399
worldsense/eQxVvndq.json#0 49 48 48.104
worldsense/mKFZqIgr.json#0 106 105 105.106

These five target end times also extend slightly beyond the probed container durations. Both the target times and recorded durations are preserved. The export does not silently clip these intervals or infer corrected labels from container metadata.

Sources & license

RVG-Bench annotations: Original question, reference-answer, temporal-interval, category, and modality annotation contributions use CC BY-NC-SA 4.0. This grant covers only the rights held by the RVG-Bench authors and applicable database rights. Attribution, noncommercial use, and share-alike conditions apply; see LICENSE.

Source videos and audio: Third-party materials retain their original licenses and applicable source terms. They are not relicensed by the RVG-Bench annotation license. Source-specific details are collected in LICENSES.md.

Upstream source Official source Published terms
FunQA Dataset card · Project CC BY-NC-SA 4.0
Daily-Omni Dataset card CC BY-NC-SA 4.0
Video-Holmes Project notice · License file Apache-2.0, with a separate academic-use statement and original video owners' copyright notice.
WorldSense Project · Dataset card CC BY 4.0 on GitHub/Hugging Face; the paper states differing terms, detailed in LICENSES.md.

The annotation license does not replace upstream terms or grant additional media rights.