README.md

---
language: en
license: mit
task_categories:
- conversational
- question-answering
tags:
- function-calling
- tool-use
- synthetic
- glm5
- agentic
size_categories:
- 1K<n<10K
---

# GLM-5.3-Flash Function Calling (synthetic)

A synthetic function-calling dataset generated with [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) via Hugging Face Inference Providers. Every example is schema-validated and deduplicated before it is written.

## Splits

| File | Format | Rows |
|---|---|---|
| `data/messages.jsonl` | OpenAI messages (multi-turn) | — |
| `data/xlam.jsonl` | xlam single-turn | — |

_The table is filled in when the generation run completes._

## `messages` format

```json
{
  "id": "glm53fc-000123",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": null, "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "...", "arguments": "{...}"}}]},
    {"role": "tool", "tool_call_id": "call_1", "name": "...", "content": "{...}"},
    {"role": "assistant", "content": "final natural-language answer"}
  ],
  "tools": [/* OpenAI function definitions used in the example */],
  "domain": "travel",
  "type": "tool" 
}
```

Multi-call examples contain several entries in one assistant `tool_calls` list plus one `tool` reply per call.

## `xlam` format

```json
{"id": "...", "query": "user request", "tools": [...], "answers": [{"name": "...", "arguments": {...}}]}
```

No-tool ("negative") examples have empty `answers` — they teach a model *not* to call a function.

## Generation pipeline

1. **Scenario** — GLM-5.3-Flash invents a domain scenario with 3–6 JSON-schema functions and a natural user query (16 domains, ~85% tool-using / 15% no-tool negatives).
2. **Call** — the same model answers with tools attached (`tool_choice="required"` / `"auto"`), producing real OpenAI-style `tool_calls`.
3. **Simulate** — the model fabricates plausible JSON tool results, then writes the final assistant answer with results in context.

Arguments are validated against each function's JSON Schema (required keys present, types match, no extra keys); failures are dropped. Generation script: [`generate_function_calling.py`](./generate_function_calling.py). Resume-safe (append + dedup), so later runs extend the dataset in place.