SciIR-82k: 80,000+ Scientific Image-Text Pairs for Reasoning Generation

SciIR-82k is a large-scale dataset for Scientific Image Reasoning Generation, containing more than 80,000 high-quality scientific image-text pairs. The samples are derived from open-access scientific publications and enriched with structured reasoning annotations. The dataset was created by author 'contton-sss' and was last updated on June 20, 2026.

Use Cases

• Training text-to-image models to generate scientifically accurate images based on structured reasoning annotations.
• Evaluating the scientific faithfulness of generated visual content from AI models.
• Researching the relationship between scientific text and corresponding visual representations.
• Developing models for automatic captioning or explanation of scientific imagery.

Strengths

• Contains more than 80,000 high-quality scientific image-text pairs.
• Samples are derived from open-access scientific publications, suggesting a credible source.
• Includes structured reasoning annotations, which likely adds depth beyond simple captions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Open-access scientific publications.

Collection Method

Derived and enriched with annotations, as described by the author.

Freshness

Last updated 2026-06-20 13:03:38; freshness should be verified.