Nemotron RL Litmus Bench v0.1: Chemical Reasoning Questions for RL Training

Litmus-Bench v0.1 is a training dataset of short-answer questions designed to improve the chemical understanding of reasoning models. The dataset was created from the ChEMBL dataset using RDKit descriptors and is released as part of NVIDIA NeMo-Gym. It was last updated on June 4, 2026.

Use Cases

• Training reinforcement learning agents on chemical property prediction based on molecular descriptors.
• Benchmarking model performance on domain-specific, short-answer question-answering tasks in chemistry.
• Improving the reasoning capabilities of language models on structured chemical data derived from ChEMBL.

Strengths

• Dataset is derived from the established ChEMBL database, a known source for bioactive molecules.
• Questions are generated using RDKit, a standard cheminformatics toolkit, suggesting a structured creation process.
• Released by NVIDIA as part of the open-source NeMo-Gym library, indicating institutional support.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Description metadata is limited; actual data quality and answer accuracy require manual inspection.

Provenance

Source

NVIDIA, derived from the ChEMBL dataset.

Collection Method

Questions were created from ChEMBL data with RDKit descriptors.

Freshness

Last updated 2026-06-04 05:01:45; freshness should be verified.