Agent behavior learning requires supervision that records not only prompts and final answers but the full interaction loop: tool calls, tool returns, environment observations, errors, retries, and verification. This release supplies frozen trajectories that capture those signals so models can learn planning, function calling, execution feedback handling, and outcome verification rather than only single-turn responses.
What Sets It Apart
- Complete trajectories with tool-level detail: each sample preserves assistant tool calls, tool returns, environment observations, and final submissions. So what: trainers can supervise fine-grained function-calling behavior and teach recovery patterns from real failure traces.
- Cross-harness sampling and turn-level masking: tasks are sampled under different tool sets and harnesses, and low-quality turns can be masked out during loss computation. So what: enables robust SFT recipes that generalize across simulated tool environments while avoiding noisy supervision.
- Execution and outcome verification retained: many samples include sandbox test results, file checks, or program tests used for verification. So what: allows constructing SFT signals tied to verifiable outcomes rather than only model-written answers.
- Diagnostic failures and recovery traces included (de-identified): not all traces are success demonstrations. So what: you can train models to detect, recover from, and report failures rather than overfit to idealized successes.
Who It's For and Trade-offs
Great fit if you want to fine-tune LLMs for agentic workflows — function calling, multi-step tool use, search planning, code execution workflows, or office/document automation — and need dense, verifiable trajectory supervision rather than isolated Q&A. It is also suited to mix‑ratio experiments comparing agent data vs. core SFT data.
Look elsewhere if you need runnable environments, live tool implementations, or replayable harnesses: the release contains static snapshots only (no tool binaries, tests, or environment servers). Also treat frozen retrieval hits and file artifacts as historical snapshots — they are not guaranteed current.
Where It Fits
Use this dataset as an L3 refined SFT slice for agent instruction‑tuning pipelines or to augment on-device models that must call functions, verify outputs, and handle execution feedback. It pairs naturally with core-domain SFT datasets when you want both knowledge/coding capability and agentic tool-use behavior in the same model.
