stindardlogic/openai-function-calling-5k

OpenAI Function Calling Format (5K)

Synthetic function-calling conversations in the OpenAI messages format (tool_calls / tool role).

Why This Dataset

Compatible with GPT-4, Mistral, Llama-3.1, Qwen2.5, and any model trained on the OpenAI chat format. Most existing function-calling datasets use abstract schemas — this uses the exact wire format models see in production.

Dataset Description

5,000 conversations across 10 tool types:

• —~80% single-turn (one tool call → result → response)
• —~20% multi-turn (sequential tool calls, e.g. search then email)

Format

Exact OpenAI messages format:

{
  "messages": [
    {"role": "system", "content": "You are a helpful AI assistant..."},
    {"role": "user", "content": "What's the weather in Tokyo?"},
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_abc123",
        "type": "function",
        "function": {"name": "get_weather", "arguments": "{"location": "Tokyo, Japan"}"}
      }]
    },
    {"role": "tool", "content": "{"temperature": 28, ...}", "tool_call_id": "call_abc123"},
    {"role": "assistant", "content": "Current weather in Tokyo: 28°C..."}
  ],
  "tools": [{"type": "function", "function": {...}}],
  "metadata": {"type": "single_turn", "tool": "get_weather"},
  "id": "abc123"
}

Use Case

• —Fine-tune LLMs for function calling (SFT on chosen response)
• —Train tool-selection and argument generation
• —Evaluate function-calling capability
• —Drop-in compatible with OpenAI fine-tuning API format (with minor adaptation)

License

Apache 2.0