| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| tags: |
| - function-calling |
| - tool-use |
| - bfcl |
| - agentic |
| pretty_name: Function-calling fine-tune data (xLAM + ToolACE + Java/JS) |
| configs: |
| - config_name: python-headline |
| data_files: train.qwen.messages.jsonl |
| - config_name: javajs-augmented |
| data_files: train.javajs.messages.jsonl |
| --- |
| |
| # Function-calling fine-tune data (xLAM + ToolACE + Java/JS) |
|
|
| Training data for fine-tuning small language models on function calling, built against the |
| BFCL non-live AST categories. OpenAI/ChatML `messages` schema; Pythonic `[func(arg=val)]` |
| labels. |
|
|
| ## Configs |
|
|
| | Config | File | Examples | Focus | |
| |---|---|---|---| |
| | `python-headline` | `train.qwen.messages.jsonl` | 140,080 | xLAM Python backbone + ToolACE + synthetic irrelevance + parallel up-weight | |
| | `javajs-augmented` | `train.javajs.messages.jsonl` | 170,077 | the above **+ 15k Java + 15k JS** (synthesized) — adds Java/JS function calling | |
|
|
| ## Schema |
|
|
| Each line: |
| `{"messages": [{"role":"system",...}, {"role":"user",...}, {"role":"assistant","content":"[func(arg=val), ...]"}]}` |
|
|
| Tools are injected into the system message as `List of tools: [...]`; the assistant label is a |
| Pythonic call list (or `[]` when no tool applies). No raw chat-template tokens (the trainer |
| applies the template); loss is on the assistant turn only. |
|
|
| ## Sources & attribution |
|
|
| - **Salesforce xLAM / APIGen** (function-calling 60k) via the `argilla/apigen-function-calling` |
| mirror — **CC-BY-4.0**. The backbone plus the synthesized irrelevance, the parallel up-weight, |
| and the Java/JS transforms all derive from it. **Attribution: Salesforce / xLAM (APIGen).** |
| - **Team-ACE/ToolACE** — **Apache-2.0**. The AST-validated Pythonic calls and the |
| refusal → `[]` (relevance-detection) slices. |
| - Deduplicated against the `gorilla-llm/Berkeley-Function-Calling-Leaderboard` test set |
| (contamination guard). |
|
|
| ## Construction |
|
|
| Built from the sources above by deterministic transformation (no LLM): conversion to Pythonic |
| call lists, hard-negative irrelevance synthesis (pairing queries with mismatched toolsets), |
| parallel-call up-weighting, and a Python → Java/JS idiom transformation (`Class.method` / |
| camelCase naming, language-specific type mapping, AST-validated). Seeded and deduplicated |
| against the BFCL test set, so the build is reproducible. |
|
|
| ## License |
|
|
| `cc-by-4.0`, governed by the xLAM backbone (attribution required); ToolACE portions are |
| Apache-2.0. Please provide per-source attribution as above. |
|
|
