What it is
18,363 scientific reasoning and tool-use SFT trajectories plus 484 RL tasks for training open agent models. Covers self-contained reasoning, Python use, and verified external lookup. 18,363 full text trajectories in 2 SFT stages plus 484 RL tasks: stage1 self-contained scientific reasoning with Python, stage2 verified tool-use from public databases, and RL prompts with reference answers and metadata.
Key results
- 10,819 stage1 reasoning trajectories
- 7,544 stage2 verified tool-use trajectories
- 484 RL tasks; Apache-2.0; Parquet
Useful for SFT and RL training of agentic LLMs that reason, call tools, and answer scientific questions.