README.md

---
license: other
pretty_name: TRACE v1.1.0
task_categories:
- visual-question-answering
- video-text-to-text
tags:
- video
- streaming-video
- temporal-reasoning
- benchmark
language:
- en
- zh
configs:
- config_name: qa
  data_files:
  - split: train
    path: data/releases/v1.1.0/qa.jsonl
- config_name: proactive
  data_files:
  - split: train
    path: data/releases/v1.1.0/proactive.jsonl
---

<div align="center">
  <h1 style="font-size: 3.5rem; line-height: 1.1; letter-spacing: 0.08em; margin-bottom: 0.35em;"><strong>TRACE</strong></h1>
  <p><strong>Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding</strong></p>
  <p>Timestamped video QA and proactive-response annotations for evaluating what a model knows, when it knows it, and how it responds.</p>
  <p>
    <a href="https://github.com/om-ai-lab/trace-bench/tree/main"><img src="https://img.shields.io/badge/GitHub-TRACE-181717?style=for-the-badge&logo=github" alt="TRACE GitHub repository"></a>
    <a href="https://arxiv.org/abs/2609.30670"><img src="https://img.shields.io/badge/arXiv-paper-6d28d9?style=for-the-badge&logo=arxiv&logoColor=white" alt="TRACE paper on arXiv"></a>
  </p>
</div>

## How TRACE works

<p align="center">
  <img src="https://huggingface.co/datasets/omlab/trace-bench/resolve/main/assets/trace-evaluation-design.png" alt="TRACE evaluation design from causal video input through the evaluation core, model adapter, model or system, and scoring" width="100%">
</p>

TRACE separates causal video delivery, model interaction, and scoring. The same
public contract makes QA and Proactive Response results auditable across models:

- **QA:** answer a question using only the video evidence available when it arrives.
- **Proactive Response:** produce the requested response inside the annotated event or state window.
- **Traceability:** retain source identities, release hashes, temporal annotations, and validation artifacts.

## Release at a glance

| Population | Records | Windows |
| --- | ---: | ---: |
| QA | 833 | — |
| Proactive Response | 415 | 1,338 |
| Proactive standard subset | 407 | 1,270 |
| Proactive sampling-stress subset | 8 | 68 |
| Tiny setup subset | 9 QA + 9 Proactive | — |

This repository contains the frozen `v1.1.0` annotation snapshot. It does not
include videos or model weights. `video_path` values are relative provenance
identifiers; users must obtain permitted source videos from their upstream
providers.

## Files

- [`data/releases/v1.1.0/qa.jsonl`](data/releases/v1.1.0/qa.jsonl): canonical QA records.
- [`data/releases/v1.1.0/proactive.jsonl`](data/releases/v1.1.0/proactive.jsonl): canonical Proactive records and response windows.
- [`data/releases/v1.1.0/manifest.json`](data/releases/v1.1.0/manifest.json): release identity, counts, and SHA-256 hashes.
- [`data/releases/README.md`](data/releases/README.md): complete file inventory and field reference.
- [`docs/dataset-card.md`](docs/dataset-card.md): populations, review coverage, and limitations.

The release also includes tiny and Proactive subset definitions, review change
and exclusion ledgers, a review summary, and a validation report. Keep these
sidecar files together when verifying the manifest.

## Data terms and limitations

This dataset contains public ground truth for evaluation, not model input.
Timestamps use the original video time coordinate and intervals are half-open.
Review coverage and known limitations are recorded in the release artifacts.

Please read [`DATA_TERMS.md`](DATA_TERMS.md) and
[`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md) before use. TRACE-owned
additions are separately described there; the MIT code license does not
relicense upstream data or videos. Upstream redistribution permissions must be
resolved before making this dataset public.

## Citation

See [`CITATION.cff`](CITATION.cff) for the TRACE citation and upstream
attribution requirements. When reporting results, include the data release,
protocol, and scorer versions used.