Motif-Technologies/agenttracer

agent-tracer SFT trajectories

⚠️ Work in progress. This dataset and AgentTracer are under active development. The schema and contents may change without notice.

1,023 multi-turn agent trajectories in OpenAI-style conversations format, with tools schemas, per-step tool calls, and per-trajectory evaluation metadata. All trajectories were generated with moonshotai/Kimi-K3.

Subset	Rows	Content
document	400	PDF / SQLite / XLSX question answering over generated documents
search	300	Multi-hop retrieval over a generated document corpus
toolcall	299	Config migration, CSV cleaning, reconciliation, log analysis
swe	24	Issue resolution in real open-source repositories

Fields

• conversations — system / user / assistant / tool messages; assistant turns may carry thinking and tool_calls.
• tools — the tool schemas available for that trajectory.
• source — {provider, name, split, subset, synthetic{model, ground_truth, system_prompt}}.
• extra — {task, sample_idx, attempt, steps, usage, missing_tools, error, duration_s, evaluation, exit_status}.
• category — agent (999) or coding (24).

Licensing

This dataset is under mixed licensing. Read this before redistributing it or training on it.

Dataset contribution — Apache-2.0. The trajectory structure, the task generators, the metadata schema, and the 999 rows of the document, search and toolcall subsets are released by the dataset authors under the Apache License 2.0. That material is procedurally generated and entirely fictional: all names, organizations, projects and documents in it are invented, and it contains no third-party or personally identifying content.

Embedded third-party source code — original licenses apply. The 24 rows of the swe subset reproduce source files from 24 real open-source repositories verbatim (~0.96 MB across 157 file paths, 87 of them complete files). That code remains under its original license — MIT (12 repositories), Apache-2.0 (9) or BSD-3-Clause (3) — and is not relicensed by this dataset. Because upstream projects carry their notices in a root LICENSE file rather than in per-file headers, most of the reproduced files carry no inline notice; the required notices are supplied instead by ATTRIBUTION.md and the verbatim license texts under LICENSES/. Retain both if you redistribute this dataset or a derivative of it.

No copyleft-licensed (GPL / LGPL / AGPL / MPL) or Creative Commons content is present anywhere in this dataset.

GitHub issue text. The swe prompts quote the body of the originating GitHub issue. Issue text is authored by individual GitHub users and licensed to GitHub under the GitHub Terms of Service; it is not covered by the repository licenses. See ATTRIBUTION.md.

Generating model. All trajectories are outputs of moonshotai/Kimi-K3, used under the Kimi K3 License. That license does not claim ownership of model outputs and does not restrict training on them, but its §2 (Model-as-a-Service operators above USD 20M revenue over 12 months) and §3 (products above 100M MAU or USD 20M monthly revenue must display "Kimi K3") may apply to what you build from this data. The §4(a) internal-use exemption does not apply once these outputs are redistributed.

If you need a build with unambiguous single-license terms, drop the 24 rows where source.subset == "swe"; the remaining 999 rows are Apache-2.0 with only the Kimi K3 conditions above attached.

Citation

@misc{agent_tracer_sft,
  title  = {agent-tracer SFT trajectories},
  year   = {2026},
  note   = {Generated with moonshotai/Kimi-K3}
}