Agent Datasets

Agent datasets train models to use tools. DeepFabric supports two modes: single-turn (one-shot tool calls) and multi-turn (extended conversations with multiple tool interactions).

Prerequisites: Agent datasets require the Spin tool service. See Tools for setup instructions.

When to Use

• Training tool-calling capabilities
• Building agents that interact with APIs or systems
• ReAct-style reasoning with action-observation loops

Single-Turn Agent

Single-turn mode generates complete tool workflows in one assistant response.

generation:
  system_prompt: "Generate tool usage examples with reasoning."
  instructions: "Create realistic scenarios requiring tools."

  conversation:
    type: chain_of_thought
    reasoning_style: agent
    agent_mode: single_turn

  tools:
    spin_endpoint: "http://localhost:3000"
    available:
      - read_file
      - write_file
      - list_files
    max_per_query: 3
    max_agent_steps: 5

output:
  system_prompt: |
    You are an AI with access to tools. Analyze tasks, execute tools, and interpret results.
  num_samples: 10
  save_as: "agent-dataset.jsonl"

Single-Turn Output

{
  "messages": [
    {"role": "system", "content": "You are an AI with access to tools."},
    {"role": "user", "content": "Find all while loops in main.py"},
    {"role": "assistant", "content": "", "tool_calls": [
      {"id": "call_0", "type": "function", "function": {"name": "search_file", "arguments": "{\"file_path\": \"main.py\", \"keyword\": \"while\"}"}}
    ]},
    {"role": "tool", "content": "[15, 42, 101]", "tool_call_id": "call_0"},
    {"role": "assistant", "content": "Found 3 while loops on lines 15, 42, and 101."}
  ],
  "reasoning": {
    "style": "agent",
    "content": [
      {"step_number": 1, "thought": "Need to search for 'while' keyword in main.py", "action": "search_file(file_path='main.py', keyword='while')"}
    ]
  },
  "tools": [...]
}

Multi-Turn Agent

Multi-turn mode creates extended conversations with multiple tool interactions, following a ReAct pattern.

generation:
  conversation:
    type: chain_of_thought
    reasoning_style: agent
    agent_mode: multi_turn
    min_turns: 2
    max_turns: 4
    min_tool_calls: 2

  tools:
    spin_endpoint: "http://localhost:3000"
    available:
      - read_file
      - write_file
      - list_files
    max_agent_steps: 5

Multi-turn datasets include: - Multiple tool call rounds - Observation-based decisions - Extended reasoning traces

Tool Options

tools:
  spin_endpoint: "http://localhost:3000"  # Spin service URL
  available:                               # Filter to specific tools
    - read_file
    - write_file
  max_per_query: 3                         # Max tools per sample
  max_agent_steps: 5                       # Max reasoning iterations
  scenario_seed:                           # Pre-populate files
    files:
      "config.json": '{"debug": true}'

CLI Usage

# Single-turn agent (requires Spin running)
deepfabric generate config.yaml \
  --conversation-type chain_of_thought \
  --reasoning-style agent \
  --agent-mode single_turn

# Multi-turn agent
deepfabric generate config.yaml \
  --agent-mode multi_turn \
  --min-turns 2 \
  --max-turns 4

Next Steps

See Tools for: - Installing and running Spin - Available VFS tools (read_file, write_file, etc.) - Creating custom tools - Mock tool execution for testing