URSA Benchmarking Sets: Single-Step Retrosynthesis Targets

URSA benchmarking sets are benchmark targets for single-step retrosynthesis from Zagribelnyy et al. (2026). The collection includes three CSV files containing product molecules for precursor prediction, with target counts ranging from 100 to 4,972. The dataset was authored by insilicomedicine and last updated on Hugging Face in May 2026.

Use Cases

• Benchmarking retrosynthesis models based on the provided product SMILES targets.
• Evaluating model performance on challenging medicinal-chemistry targets using the expert benchmark subset.
• Testing model scalability on a curated test split from the USPTO-50K dataset.
• Conducting rapid model validation using a fixed random subsample of the main test set.

Strengths

• Includes an expert-curated benchmark of 100 challenging medicinal-chemistry targets.
• Provides a large, curated test split of 4,972 targets from the USPTO-50K dataset.
• Offers a fixed random 10% subsample (497 targets) for faster iteration and validation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

insilicomedicine on Hugging Face, based on work by Zagribelnyy et al.

Collection Method

Likely curated from existing chemical reaction datasets and expert knowledge.

Freshness

Last updated 2026-05-28 07:26:55; freshness should be verified.