Chemistry Bench

The model writes the rest of the record it thinks this structure belongs to, once per version. Versions may choose different kinds of record, which is how it shows what it is unsure about. Randomness 1.0 is what the scored reading used; lower values show its single most likely record.

Examples come from data held out of training, with their measured or calculated values, which appear beside the model's answer.

A polymer cannot be aimed at a Tg: every one of the 1,859,180 polymer records the model read puts the structure first, so it has never seen that question.

Trained on reactions from US patents (via ORDerly). Reactions unlike patent chemistry are outside what it read.

Examples are held-out patent reactions; the product and yield the patent records appear beside the model's.