SFT Records

SFT Records

Supervised fine-tuning (SFT) records teach the model to produce desired outputs for given inputs.

Requirements

• At least one assistant message in the messages array
• Messages should represent a complete interaction

Basic Example

{
  "id": "sft-001",
  "messages": [
    {
      "role": "user",
      "content": [{ "type": "text", "text": "What is the capital of Japan?" }]
    },
    {
      "role": "assistant",
      "content": [{ "type": "text", "text": "The capital of Japan is Tokyo." }]
    }
  ]
}

Multi-Turn Conversations

SFT records support multi-turn interactions:

{
  "id": "sft-multi-001",
  "messages": [
    {
      "role": "system",
      "content": [{ "type": "text", "text": "You are a helpful coding assistant." }]
    },
    {
      "role": "user",
      "content": [{ "type": "text", "text": "Write a Python function to reverse a string." }]
    },
    {
      "role": "assistant",
      "content": [{ "type": "text", "text": "def reverse_string(s):\n    return s[::-1]" }]
    },
    {
      "role": "user",
      "content": [{ "type": "text", "text": "Can you add a docstring?" }]
    },
    {
      "role": "assistant",
      "content": [{ "type": "text", "text": "def reverse_string(s):\n    \"\"\"Reverse a string.\n    \n    Args:\n        s: The string to reverse.\n    \n    Returns:\n        The reversed string.\n    \"\"\"\n    return s[::-1]" }]
    }
  ]
}

With Tool Use

SFT records can include tool calls:

{
  "id": "sft-tool-001",
  "messages": [
    {
      "role": "user",
      "content": [{ "type": "text", "text": "What is 15% of 240?" }]
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "tool_call",
          "name": "calculator",
          "call_id": "calc-1",
          "arguments": { "expression": "240 * 0.15" }
        }
      ]
    },
    {
      "role": "tool",
      "content": [
        {
          "type": "tool_result",
          "name": "calculator",
          "call_id": "calc-1",
          "result": { "value": 36 }
        }
      ]
    },
    {
      "role": "assistant",
      "content": [{ "type": "text", "text": "15% of 240 is 36." }]
    }
  ]
}

With Reasoning

Include reasoning traces for chain-of-thought training:

{
  "id": "sft-reasoning-001",
  "messages": [
    {
      "role": "user",
      "content": [{ "type": "text", "text": "If a train travels 120 miles in 2 hours, what is its average speed?" }]
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "reasoning",
          "text": "To find average speed, I divide total distance by total time: 120 miles / 2 hours = 60 miles per hour."
        },
        {
          "type": "text",
          "text": "The train's average speed is 60 miles per hour."
        }
      ]
    }
  ]
}

Multimodal SFT

Include images, audio, or other media:

{
  "id": "sft-image-001",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "What do you see in this image?" },
        { "type": "image", "ref": { "asset_id": "img-001" } }
      ]
    },
    {
      "role": "assistant",
      "content": [{ "type": "text", "text": "The image shows a sunset over the ocean with vibrant orange and purple colors." }]
    }
  ]
}

Best Practices

1. Vary conversation lengths - Include short and long conversations
2. Include system messages - Or use defaults for consistency
3. Quality over quantity - High-quality examples beat noisy data
4. Cover edge cases - Include refusals, corrections, and clarifications
5. Balance topics - Avoid over-representing any single domain