opendatalab/MolRecBench-Wild

MolRecBench-Wild (2026-08-19)

MolRecBench-Wild is a real-world benchmark for optical chemical structure recognition. This release contains 5,024 molecular structure images and their CARBON molecular-graph annotations from 818 source articles.

This is the repository's authoritative 2026-08-19 release. It differs from the 5,029-sample snapshot described in the first arXiv version of the paper.

Dataset structure

The dataset has one test split. Images are embedded in native Parquet files, so the Hugging Face Dataset Viewer can display each image next to its structured annotation.

from datasets import load_dataset

dataset = load_dataset("opendatalab/MolRecBench-Wild", split="test")
sample = dataset[0]
sample["image"].show()
print(sample["id"], sample["symbols"], sample["bonds"])

Fields

• —image: cropped molecular structure image.
• —id: stable sample identifier and original image filename.
• —release_id: dataset release date (2026-08-19, ISO 8601 format).
• —source, source_doi, source_url: provenance derived from the DOI encoded in the sample ID.
• —evaluation_subset: benchmark difficulty subset (A, B, or C).
• —hardcase_label: visual and chemical difficulty labels.
• —symbols, charges, radicals, valences, isotopes, attach_points, coords, bonds, brackets: CARBON graph annotation fields.

Subset sizes are 1,987 A samples, 1,976 B samples, and 1,061 C samples. Subset A has no specified chemical-semantic property difficulty labels and relatively few visual difficulty labels; B has no such property labels but more visual difficulty labels; C contains specified chemical-semantic property difficulty labels.

Source information

A DOI and resolver URL are included with every sample to identify its source article. This Dataset Card does not declare a license for this release.

Citation

Please cite the MolRecBench-Wild paper:

• —MolRecBench-Wild: A Real-World Benchmark for Optical Chemical Structure Recognition, CVPR 2026, arXiv:2605.05832.

Release integrity

Checksums, row counts, Parquet shard metadata, and the image-set fingerprint are available in release_manifest.json.