Motif-Technologies/agenttracer-bfcl

agent-tracer BFCL multi-turn trajectories

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

2,310 multi-turn function-calling trajectories in OpenAI-style conversations format, with tools schemas, per-step tool calls, and per-trajectory evaluation metadata. The record layout follows Motif-Technologies/agenttracer.

Each trajectory is a stateful, multi-turn task over the executable API classes of the Berkeley Function Calling Leaderboard (BFCL v4, bfcl-eval==2026.3.23): file system, messaging, posting, tickets, trading, travel, vehicle control and math. Tasks were newly written for this dataset (no BFCL evaluation question, initial state or answer was used as a seed), every call was executed against the real API classes, and only trajectories that passed the upstream BFCL multi_turn_checker (plus stricter state checks on turns whose ground truth is empty) were kept.

Variant	Rows	Content
miss_func	1,259	A function needed in one turn is withheld; the assistant must say it cannot do it without changing state, then finish once the function is added
base	558	Multi-turn requests completed with the provided functions
long_context	493	Same as base, executed with BFCL's long-context environment state

All rows have source.subset == "toolcall"; the variant is not stored per row.

Teacher model	Rows
z-ai/glm-5.3-flash	2,244
moonshotai/Kimi-K3	66

Format notes

• No system prompt, no submit / code_exec tools: each user turn ends with an assistant reply that makes no tool call, as in BFCL's FC evaluation. extra.exit_status is completed.
• Assistant turns carry the teacher's thinking. Turns where the model reported 0 reasoning tokens have an empty thinking; trajectories with no reasoning at all were rejected.
• Tool descriptions carry BFCL's Python-syntax hint ("Note that the provided function is in Python 3 syntax.").
• Records of 32,768 tokens or more (Qwen3.5 chat template, reasoning kept on every turn) were removed.
• Tasks whose user turns share an 8-gram with any BFCL multi-turn evaluation question were removed.

Fields

• conversations — 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{synth, model, effort, ground_truth, system_prompt}}.
• extra — {task, sample_idx, attempt, steps, missing_tools, error, evaluation, exit_status}.
• category — agent (2,310).

Licensing

Dataset contribution — Apache-2.0. The task texts, initial environment states and the trajectory structure are released under the Apache License 2.0. They are procedurally generated and fictional: names, accounts, files and messages are invented.

BFCL environment and tool schemas — Apache-2.0. The tool schemas and the executable API classes come from bfcl-eval (Berkeley Function Calling Leaderboard, gorilla), licensed under Apache-2.0. Keep that attribution if you redistribute this dataset.

Generating models. 2,244 trajectories are outputs of z-ai/glm-5.3-flash, which is released under the MIT License. 66 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. If you need outputs from MIT-licensed models only, drop the rows where source.synthetic.model == "moonshotai/Kimi-K3".

Citation

@misc{agent_tracer_bfcl,
  title  = {agent-tracer BFCL multi-turn trajectories},
  year   = {2026},
  note   = {Generated with z-ai/glm-5.3-flash and moonshotai/Kimi-K3 on BFCL v4 executable APIs}
}