Zebra-CoT: 182,384 Interleaved Vision-Language Reasoning Traces

Zebra-CoT is a large-scale dataset containing 182,384 samples of logically coherent interleaved text and image reasoning traces. It was created by multimodal-reasoning-lab and covers four major categories: scientific reasoning, 2D visual reasoning, 3D visual reasoning, and visual logic & strategic games. The dataset was last updated on Hugging Face in January 2026.

Use Cases

• Training multimodal large language models (MLLMs) based on the interleaved reasoning traces.
• Benchmarking model performance on scientific reasoning tasks described in the dataset.
• Evaluating visual logic and strategic game-solving capabilities using the provided examples.
• Developing models for 2D and 3D visual reasoning based on the dataset's structured problems.

Strengths

• Contains 182,384 samples, indicating a substantial scale for training and evaluation.
• Covers four distinct reasoning categories, suggesting diversity in problem types.
• Each example includes a logically coherent interleaved text-image reasoning trace.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific file formats, sample data, and license information are unknown.
• Data may reflect bias inherent to the sources and methods used for its creation.

Provenance

Source

multimodal-reasoning-lab on Hugging Face

Collection Method

Likely curated or generated for multimodal reasoning research.

Time Range

null

Freshness

Last updated 2026-01-30 22:59:20; freshness should be verified.

Geography

null