README.md

metadata

pretty_name: Mat-20K-RL
language:
  - en
task_categories:
  - text-generation

Mat-20K-RL

Mat-20K-RL contains 20,000 materials-science queries with reference solutions for reinforcement-learning and verifier development. It includes 10,000 crystal-structure-design queries and 10,000 materials-property-prediction queries sampled from the Mat-252K-SFT training set.

Each record contains:

• mp_id: Materials Project identifier.
• question_type: structure_design or property_prediction.
• messages: the sampled user query in chat format.
• solution: the reference CIF or property JSON associated with mp_id.

The dataset license and archival DOI will be added before public release.