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. ```python 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](https://arxiv.org/abs/2605.05832). ## Release integrity Checksums, row counts, Parquet shard metadata, and the image-set fingerprint are available in `release_manifest.json`.