README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - 1K<n<10K
task_categories:
  - text-generation
pretty_name: Agentic DPO Tool-Use Pairs (3K)
tags:
  - dpo
  - preference-pairs
  - tool-use
  - agentic
  - rlhf
  - function-calling
  - alignment
  - synthetic
configs:
  - config_name: default
    data_files:
      - split: train
        path: agentic-dpo-tool-use-3k.jsonl

Agentic DPO Tool-Use Pairs (3K)

Synthetic DPO preference pairs for training LLMs to use tools correctly in agentic settings.

Dataset Description

3,000 preference pairs covering 7 tool categories:

• web_search — real-time web search
• calculator — mathematical expression evaluation
• weather_api — current weather retrieval
• code_interpreter — Python code execution
• database_query — SQL database queries
• stock_price — financial data lookup
• translate — multilingual translation

Each example contains:

• system: system prompt with tool definitions
• prompt: user query requiring tool use
• chosen: correct response that uses the appropriate tool and interprets results
• rejected: flawed response (refuses tool use / hallucinates / ignores tool / claims no access)
• metadata: category and contextual info

Format

{
  "system": "You are a helpful AI assistant with access to the following tools: ...",
  "prompt": "What's the current price of NVDA?",
  "chosen": "<tool_call>{...}</tool_call>\n<tool_result>{...}</tool_result>\n**NVDA** is trading at ...",
  "rejected": "NVDA stock is around $875 but I don't have real-time data...",
  "metadata": {"category": "stock_price", "ticker": "NVDA"},
  "id": "abc123def456"
}

Use Case

Fine-tune models to:

1. Recognize when a tool is needed vs. when to answer from knowledge
2. Call tools correctly with proper parameters
3. Interpret and present tool results clearly

Compatible with DPO, IPO, KTO, and other preference optimization frameworks.

License

Apache 2.0