MolParser-7M: 8 Million Molecule Image-SMILES Pairs for OCSR

MolParser-7M contains approximately 8 million paired molecule images and chemical structure strings, developed by UniParser for the ICCV 2025 paper "MolParser: End-to-end Visual Recognition of Molecule Structures in the Wild." The collection features a pre-training subset of 7.7 million synthetic images paired with a specialized extended-SMILES format.

Use Cases

• Training end-to-end visual recognition models for molecule structures using the extended-SMILES labels
• Evaluating optical chemical structure recognition (OCSR) performance on molecule images
• Fine-tuning vision-language models for chemical informatics and structure-to-text tasks

Strengths

• Scale of 7.7 million synthetic training records
• Peer-reviewed origin from ICCV 2025
• Includes a specialized extended-SMILES format for structural representation

Limitations

• The 7.7M pre-training subset is entirely synthetic data
• Restricted to non-commercial use under CC BY-NC-SA 4.0 license
• Specific column names are not documented in the metadata

Provenance

Source

UniParser, "MolParser: End-to-end Visual Recognition of Molecule Structures in the Wild" (ICCV 2025)

Collection Method

Synthetic generation for the 7.7M pre-training subset

Freshness

Last updated December 2025; corresponds to ICCV 2025 research findings.