| --- |
| language: |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - function-calling |
| - tool-use |
| - synthetic |
| - glm |
| - conversational |
| size_categories: |
| - n<1K |
| --- |
| |
| # GLM-5.3-Flash function-calling (synthetic) |
|
|
| 500 synthetic function-calling training samples generated with |
| [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) via HF Inference |
| Providers (auto-routing with novita / together / fireworks fallback), 2026-09-21. |
|
|
| ## Schema |
|
|
| Each row: |
|
|
| - `messages` — chat in OpenAI tool-use format (`system` / `user` / `assistant`); the final |
| assistant message either carries `tool_calls` (with JSON-string `arguments`) or is a plain-text answer |
| - `tools` — 1–4 tool schemas in OpenAI function format (`type`, `function{name, description, parameters}`) |
| - `category` — one of: |
| - `single_call` — exactly one tool call |
| - `parallel_calls` — 2–4 independent calls in one response |
| - `multi_turn` — multi-turn with tool-result turns |
| - `error_handling` — call with plausibly failing arguments |
| - `no_tool` — tempting but irrelevant tools; assistant answers directly |
|
|
| Domains span weather, travel, finance, e-commerce, smart home, file management, and ~14 more. |
| Every sample passed a validator: tool-call names resolve to declared tools, `arguments` parse as |
| JSON objects, last message is assistant-role, message roles valid. |
|
|
| ## Generation |
|
|
| Single-shot JSON generation per sample (`temperature=0.9`, seeded), 10-way concurrency, |
| per-sample validation with provider fallback. Generation script pattern available in the |
| conversation that produced this dataset. Costs: ~7 min on `cpu-basic` (Jobs) + provider inference. |
