CoolFace
Datasetpublic

Keylab/OCSR-Benchmarks

OCSR Benchmarks A collection of ten benchmark datasets for Optical Chemical Structure Recognition (OCSR) — the task of converting chemical structure diagram images into machine-readable SMILES strings. These benchmarks were used to evaluate the COMO model (Closed-Loop Optical Molecule Recognition). Subsets Config Split Size Domain CLEF test 992 Real JPO test 449 Real UOB test 5,740 Real USPTO test 5719 Real USPTO-10K test 9,999 Real Staker… See the full description on the dataset page: https://huggingface.co/datasets/Keylab/OCSR-Benchmarks.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
3likes355downloads
Dataset Card

OCSR Benchmarks

A collection of ten benchmark datasets for Optical Chemical Structure Recognition (OCSR) — the task of converting chemical structure diagram images into machine-readable SMILES strings.

These benchmarks were used to evaluate the COMO model (Closed-Loop Optical Molecule Recognition).

Subsets

ConfigSplitSizeDomain
CLEFtest992Real
JPOtest449Real
UOBtest5,740Real
USPTOtest5719Real
USPTO-10Ktest9,999Real
Stakertest50,000Real
ACStest331Real
WildMol-10Ktest9,889Real
Indigotest5,719Synthetic
ChemDrawtest5,719Synthetic

Schema

Each sample has three fields:

FieldTypeDescription
image_idstringOriginal identifier for the sample
imageImagePNG image of the chemical structure diagram
SMILESstringGround-truth SMILES string

Usage

python
from datasets import load_dataset

# Load a single benchmark
ds = load_dataset("Keylab/OCSR-Benchmarks", name="USPTO", split="test")
sample = ds[0]
sample["image"].show()   # PIL Image
print(sample["SMILES"])

# Iterate over all benchmarks
for config in ["CLEF", "JPO", "UOB", "USPTO", "USPTO-10K",
               "Staker", "ACS", "WildMol-10K", "Indigo", "ChemDraw"]:
    ds = load_dataset("Keylab/OCSR-Benchmarks", name=config, split="test")
    print(f"{config}: {len(ds)} samples")

Bulk Download

Pre-packaged .tar.gz archives (images + CSV) are also available in the COMO model repository for direct download without the datasets library.

License

These benchmarks are collected from existing public OCSR datasets. Please refer to the original sources for attribution and applicable terms:

DatasetSource
USPTO, CLEF, JPO, UOB, StakerRajan et al., 2020, Xiong et al., 2023
Indigo, ChemDraw, ACS, StakerQian et al., 2023
USPTO-10KMorin et al., 2023
WildMol-10KFang et al., 2025

Citation

If you use these benchmarks, please cite the COMO paper and the original benchmark sources:

bibtex
@article{lyu2026closed,
  title={COMO: Closed-Loop Optical Molecule Recognition with Minimum Risk Training},
  author={Lyu, Zhuoqi and Ke, Qing},
  journal={arXiv preprint arXiv:2604.23546},
  year={2026}
}