Audit finds MoLFormer's stereochemistry claim is wrong about pretraining
- The MoLFormer model card says isomeric information was stripped from training data, yet the repository's own token-frequency table shows stereocentre tokens are common.
- Stripping does exist, but it sits in fine-tuning scripts that hardcode non-isomeric SMILES for downstream benchmarks, not in pretraining.
- SMI-TED's released vocabulary size differs from its reported one, and its frequency table matches MoLFormer's despite a described different corpus.
- The authors share the extracted table and an audit tool, and urge model cards to state stereochemistry handling per stage.
💡 Points Based on the abstract
Before adopting any off-the-shelf embedding, I'd open the tokenizer and the fine-tuning scripts, not the model card. If the downstream benchmarks were run on non-isomeric SMILES, they can't show whether the model resolves chirality, so reported gains say nothing about stereo-driven SAR. I'd add enantiomer and diastereomer pair tests to our own evaluation first.
📄 Paper · Journal of Cheminformatics ↗ Model cards are not primary artifacts: stereochemistry in the MoLFormer and SMI-TED chemical foundation models