ChemCensor Database

This repository provides the SQLite databases used by ChemCensor, a precedent-based framework for evaluating reaction chemical plausibility.

ChemCensor separates the reaction center (what changes) from the functional-group context (what must be tolerated), then checks whether similar patterns are supported by documented synthetic precedents. The resulting score ranges from 0 to 5, where higher values indicate stronger precedent support.

Database versions and compatibility

Database	Compatible ChemCensor versions	Status
ChemCensor-DB-U3.sqlite	>=1.3.0	Current
ChemCensor-DB-U2-1.0.0.sqlite.zip	>=1.0.0,<1.3.0	Legacy

The current U3 database uses the stereo-aware RC1 key format introduced in ChemCensor 1.3.0. It must not be used with ChemCensor 1.2.x or older. The U2 archive remains available for those older package versions.

Automatic download

ChemCensor 1.4.0 and newer download and cache the current database automatically on first use:

from chemcensor import ChemCensor

censor = ChemCensor.open()
score = censor.score("CCO>>CC=O")

The file is stored in the standard Hugging Face cache and reused on subsequent calls.

To download explicitly from Python:

from chemcensor import download_default_database

db_path = download_default_database()

Or use the Hugging Face CLI:

hf download insilicomedicine/chemcensor \
  ChemCensor-DB-U3.sqlite \
  --repo-type dataset \
  --local-dir data

Then open the local file:

from chemcensor import ChemCensor

censor = ChemCensor.open("data/ChemCensor-DB-U3.sqlite")

What is inside

The knowledge base was constructed from the USPTO-full reaction dataset introduced by D. Lowe and includes reactions extracted by text mining from United States patents published between 1976 and September 2016.

The database includes:

• reaction-center precedent information;
• functional-group compatibility information;
• relational mappings required by the ChemCensor scoring pipeline.

Citation

If you use this database or ChemCensor in your work, please cite:

@misc{zagribelnyy2026singleanswerenoughrethinking,
      title={When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs},
      author={Bogdan Zagribelnyy and Ivan Ilin and Maksim Kuznetsov and Nikita Bondarev and Roman Schutski
                and Thomas MacDougall and Rim Shayakhmetov and Zulfat Miftakhutdinov
                and Mikolaj Mizera and Vladimir Aladinskiy and Alex Aliper and Alex Zhavoronkov},
      year={2026},
      eprint={2602.03554},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.03554},
}

License

This database is part of ChemCensor and is distributed under the same terms as the ChemCensor software. By downloading or using this dataset, you agree to the ChemCensor License Agreement (independent benchmarking and evaluation purposes only).

Use in products, automated pipelines, training or fine-tuning of models, redistribution, or incorporation into operational workflows is not permitted without prior written permission from Insilico Medicine AI Limited.