Tool Use / Function Calling Training

In one sentence: TODO — teach the model to issue tool calls at the right time and with the correct schema, and to digest the tool's returned results.

Prerequisites: SFT Overview, Chat Template

Status

🚧 This page is a placeholder outline; the full text has not been written yet.

1. Intuition and Motivation

TODO:

• [ ] Knowledge inside model weights is limited: computation, retrieval, and acting on the external world require tools
• [ ] A tool call is essentially constrained structured generation

2. Data and Format

TODO:

• [ ] How tool schemas are injected into the prompt (JSON schema in the system section)
• [ ] Role design for call turns / tool-result turns in the chat template
• [ ] Data synthesis pipeline: API library → scenario generation → trajectory generation → validation and filtering (the ToolLLM / APIGen approach)

3. Training Methods

TODO:

• [ ] SFT: compute loss on tool-call turns, mask out tool-result turns
• [ ] Preference optimization: construct DPO pairs from correct vs incorrect calls
• [ ] When to decline a call / answer directly (negative samples)

4. Evaluation

TODO: BFCL, τ-bench, etc.

5. Experiments and Tuning Experience

TODO: common failures: hallucinated arguments, missed calls, excessive calls.

6. References

• [ ] Qin et al., 2023. ToolLLM
• [ ] Schick et al., 2023. Toolformer