Reproducing an agent model usually means reconstructing not just its weights, but also the data and interaction traces that taught it to reason and use tools. This release makes those traces available: 18,363 supervised fine-tuning conversations and 484 reinforcement-learning tasks from the Axion 1.0 training recipe. The useful distinction is between learning to solve a problem with what is already given and learning to retrieve missing evidence before answering.
What the Data Makes Possible
- Stage 1 contains 10,819 self-contained scientific and mathematical reasoning trajectories; Python execution is allowed, but retrieval is not. This isolates reasoning from search.
- Stage 2 contains 7,544 verified tool-use trajectories built around public structured data. The conversation records include tool calls and responses, so researchers can examine the interaction rather than only the final answer.
- The separate RL split contains prompts, reference answers, and task metadata—not completed solution conversations. Keeping this distinction clear matters when selecting data for supervised versus reward-based training.
- The complete recipe is designed to support reproducibility and extension, rather than offering weights alone.
Who Should Use It
A good fit for researchers studying agent training, scientific reasoning, or tool-mediated problem solving who want inspectable training examples and a starting point for reproducing the Axion recipe. Look elsewhere if you need general-purpose chat data, broad coverage of everyday tasks, or fully completed RL trajectories: this collection is focused on scientific reasoning and structured, verifiable tool tasks.
