Many chemistry reasoning tasks are hard to use for reinforcement learning because judging an answer can require specialist review. This collection makes a broad set of chemistry questions short-answer and automatically checkable, connecting domain knowledge with verifiable-reward post-training.
What Sets It Apart
- Its 47,118 training and 2,056 test questions span 126 properties and 12 assay endpoints, giving training a wider target range than a single-property benchmark.
- Questions draw on ChEMBL, ExpansionRx, and Biogen data, with some properties derived using RDKit. This combines molecular and assay-related tasks rather than relying on one source or descriptor type.
- The short-answer format is designed for verifiable rewards in reinforcement learning, so model outputs can be scored without relying solely on open-ended human judgment.
- Licensing is mixed: ChEMBL-derived questions use CC BY-SA 4.0, while other questions use OpenMDW 1.1. Review the applicable terms before redistributing or using the data commercially.
Who It’s For—and the Trade-Off
A good fit for teams post-training language models on chemistry reasoning or testing whether they can answer structured molecular questions. Look elsewhere for a general-purpose chemistry corpus, a standalone molecular-property prediction benchmark, or tasks that demand detailed explanations: the emphasis here is on short, checkable answers.
