stindardlogic/tool-calling-english-100k

Tool Calling English (100K)

100,000 tool-calling conversations in OpenAI function calling format — the largest general English tool-use dataset for fine-tuning.

Motivation

Models trained without tool-calling examples struggle in agentic deployments. This dataset trains the full cycle: deciding when to call a tool, calling it with correct arguments, interpreting the result, and producing a grounded final response.

Dataset Description

100,000 conversations across 19 real-world use case categories and 20 tool schemas:

Tool Schemas Included

Tool	Description
get_weather	Current conditions + multi-day forecast
search_web	Web search with date filtering
execute_python	Sandboxed Python execution
query_database	SQL read queries
send_email	Email with CC and priority
get_stock_price	Real-time stock data
create_calendar_event	Calendar scheduling with invites
read_file / write_file	File system operations
translate_text	Multi-language translation
get_exchange_rate	Foreign exchange rates
search_flights	Flight search with cabin class
get_directions	Navigation with transport mode
create_chart	Data visualization
send_slack_message	Slack messaging with threads
list_files	Directory listing
get_news	News by category
calculate	Math expression evaluation
get_repository_info	GitHub repo metadata
search_products	E-commerce product search

Dataset Composition

Characteristic	Count	%
Single tool call	~78,000	78%
Parallel tool calls	~13,000	13%
Multi-turn conversations	~8,700	9%
No tool needed (model declines correctly)	~8,600	9%

Categories

Category	Examples
Finance (stock + exchange)	~13,200
Travel (flights + weather)	~8,800
Weather	~8,600
Code execution	~4,500
Calculation	~4,400
Web search	~4,400
Knowledge Q&A (no tool)	~4,400
Navigation	~4,400
Data analysis	~4,400
Calendar, Email, Database, Productivity, Translation, Files, Shopping, Dev tools, News	~4,200 each

Format

Standard OpenAI function calling format:

{
  "messages": [
    {"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", "units": "celsius"}"}}]
    },
    {"role": "tool", "tool_call_id": "call_abc123", "content": "{"temperature": 22, "condition": "Partly cloudy"}"},
    {"role": "assistant", "content": "The current weather in Tokyo is 22°C and partly cloudy..."}
  ],
  "tools": [{"type": "function", "function": {"name": "get_weather", ...}}],
  "metadata": {"category": "weather", "has_tool_call": true, "parallel_calls": false, "multiturn": false},
  "id": "abc123"
}

Key Training Signals

1. When to call — 8.6% of conversations have no tool call because the model knows the answer (prevents over-reliance on tools)
2. Correct arguments — arguments match the schema exactly, including optional parameters and enum values
3. Parallel calls — 13% demonstrate calling multiple tools simultaneously when independent
4. Result grounding — final responses reference the tool output, not prior knowledge
5. Multi-turn — 8.7% show sequential tool use across conversation turns

Training Compatibility

# LLaMA-Factory
# Set dataset_format: openai in config

# Direct loading
import json
with open("tool-calling-english-100k.jsonl") as f:
    data = [json.loads(line) for line in f]

# Filter by characteristic
parallel = [r for r in data if r["metadata"]["parallel_calls"]]
no_tool = [r for r in data if r["metadata"]["no_tool_needed"]]

License

Apache 2.0