README.md

metadata

license: cc-by-4.0
task_categories:
  - video-text-to-text

SYNCR

Paper | GitHub

SYNCR is a synthetic cross-video reasoning benchmark for evaluating multimodal large language models on questions that require reasoning across multiple independent videos.

The benchmark contains 8,163 multiple-choice question-answer examples spanning 4 reasoning categories and 8 tasks. Each example is programmatically grounded in synthetic video data generated from CLEVRER, Kubric, and Habitat-based environments.

Dataset contents

The Hugging Face dataset viewer exposes the annotation file:

• test/metadata.jsonl

Large video assets are provided in this repository as archive files and/or task-specific folders. The videos field in each annotation row points to the video files associated with that question.

Reasoning categories and tasks

SYNCR covers four reasoning categories:

1. Temporal Alignment

   • CLEVRER temporal ordering
   • Kubric synchronization

2. Spatial Tracking

   • Kubric spatial tracking
   • Habitat agent/object tracking

3. Comparative Reasoning

   • CLEVRER kinematic comparison
   • CLEVRER collision comparison

4. Holistic Synthesis

   • Habitat object counting
   • Habitat route planning

Sample usage

To generate SYNCR benchmark data, use the unified generate_data.py script provided in the GitHub repository:

# Example: Generate Temporal Alignment data (Multi-angle synchronization)
python generate_data.py --dataset kubric_sync --path /path/to/kubric --total-num 1000 --save

# Example: Generate Temporal Alignment data (Sequential ordering)
python generate_data.py --dataset clevrer_temporal --root-path /path/to/clevrer --total-num 1000 --save

Splits

The current release provides one split:

• test: 8,163 examples

Schema

Each row contains the following fields:

• category: the high-level reasoning category
• task: the specific benchmark task
• question: the natural-language multiple-choice question
• options: the answer choices
• answer: the correct answer
• videos: the list of video paths associated with the question
• video_ranges: optional temporal ranges or clip spans associated with the videos

Intended use

SYNCR is intended for evaluating cross-video reasoning in multimodal models. It is designed for diagnostic evaluation rather than model training, with programmatically generated questions and answers that support controlled analysis of spatial, temporal, comparative, and holistic reasoning failures.

License

The SYNCR dataset (including video assets and annotations) is provided under the Creative Commons Attribution 4.0 International (CC-BY 4.0) License.

The code for benchmark generation and evaluation is provided separately under the MIT License in the accompanying GitHub repository.