高品質SFTデータ構造設計のファインチューニング

高品質SFTデータ構造

Posted by LuochuanAD on March 15, 2026 本文总阅读量

背景

ファインチューニングモデルの性能 = 70% データ構造設計 + 20% データ品質 + 10% トレーニングパラメータ

この記事は「データ構造設計」にのみ焦点を当てます。

高品質SFTデータの5つのレベル

  1. Level 1 基本的なQ&A
  2. Level 2 システムプロンプト
  3. Level 3 思考の連鎖 (CoT)
  4. Level 4 多段対話
  5. Level 5 ツール / RAG

分析

Level 1 基本的なQ&A

基礎的な応答能力であり、誤答や推論の不足がある

{
 "messages":[
  {"role":"user","content":"What is XAI?"},
  {"role":"assistant","content":"XAI is the process..."}
 ]
}

Level 2 システムプロンプト

systemが追加され、モデルに役割行動を学習させ人格を形成させる

{
 "messages":[
  {
   "role":"system",
   "content":"You are a professional biology teacher."
  },
  {
   "role":"user",
   "content":"Explain photosynthesis"
  },
  {
   "role":"assistant",
   "content":"Photosynthesis is..."
  }
 ]
}

Level 3 思考の連鎖 (CoT)

推論能力の向上

{
 "messages":[
  {"role":"user","content":"If I have 20 apples and give away 7, how many remain?"},
  {"role":"assistant","content":
   "Let's think step by step.\n
    1. Start with 20 apples\n
    2. Give away 7 apples\n
    3. 20 - 7 = 13\n
    Therefore, 13 apples remain."
  }
 ]
}

Level 4 多段対話

実際の会話は多段のやり取りであり、モデルはコンテキストの記憶と連続推論を学習する

{
 "messages":[
  {"role":"system","content":"You are a math tutor."},
  {"role":"user","content":"What is 12 * 8?"},
  {"role":"assistant","content":"12 × 8 = 96."},
  {"role":"user","content":"What about 96 ÷ 3?"},
  {"role":"assistant","content":"96 ÷ 3 = 32."}
 ]
}

Level 5 ツール / RAG

モデルはユーザーの質問に応じてツールの選択、API呼び出し、応答返却を学習する

{
 "messages":[
  {"role":"user","content":"What's the weather in Tokyo?"},
  {
   "role":"assistant",
   "content":"I will check the weather API.",
   "tool_call":{
      "name":"get_weather",
      "arguments":{"city":"Tokyo"}
   }
  },
  {
   "role":"tool",
   "content":"Temperature: 20°C"
  },
  {
   "role":"assistant",
   "content":"The temperature in Tokyo is 20°C."
  }
 ]
}

モデルはコンテキストを読み込む

{
 "messages":[
  {
   "role":"system",
   "content":"Use the provided context to answer."
  },
  {
   "role":"user",
   "content":"Context: Tokyo is the capital of Japan.\nQuestion: What is the capital of Japan?"
  },
  {
   "role":"assistant",
   "content":"The capital of Japan is Tokyo."
  }
 ]
}

推奨される高品質SFTデータ構造例

{
 "messages":[
  {"role":"system","content":"You are a helpful AI."},

  {"role":"user","content":"Calculate 25 * 14"},

  {"role":"assistant","content":
  "Let's solve step by step.\n
   25 × 14 = (25 × 10) + (25 × 4)\n
   = 250 + 100\n
   = 350\n
   Therefore the answer is 350."
  }
 ]
}