ChemCoT: 10K-100K Records for Molecular Editing and Reaction Reasoning

ChemCoT provides between 10,000 and 100,000 text-based records for chemical reasoning tasks, released by IDEA-AI4S in 2025. The collection is organized into four modules covering molecular editing, property optimization, structural understanding, and chemical reactions to support Chain-of-Thought modeling.

Use Cases

• Predicting chemical reaction outcomes and retrosynthetic routes using the rxn module files like retro.json and fs_major_product.json
• Optimizing molecular properties such as solubility and lipophilicity using the mol_opt files solubility.json and logp.json
• Analyzing molecular topology and functional groups using ring_count.json and fg_count.json within the mol_und module

Strengths

• Contains 10,000 to 100,000 records
• Covers 20+ distinct chemical task categories
• MIT licensed for open research
• Includes specialized reaction data like retrosynthesis and mechanism selection

Limitations

• Text-only modality lacks 3D molecular coordinates
• Specific internal schema for JSON files is not documented in the summary
• Geographic bias toward US-based regions as per metadata tags

Provenance

Source

IDEA-AI4S

Freshness

Last updated February 2026

Geography

United States