[{"data":1,"prerenderedAt":212},["ShallowReactive",2],{"topic-page:\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Ffoundation-model-training-fine-tuning-and-data":3},{"topic":4,"category":171,"seo":175,"breadcrumbs":180,"prerequisites":190,"nextTopics":191,"relatedTopics":200},{"id":5,"slug":6,"title":7,"summary":8,"canonicalPath":9,"categoryPath":10,"difficulty":11,"targetAudience":12,"estimatedMinutes":13,"generatedAt":14,"publishedAt":14,"updatedAt":14,"reviewStatus":15,"tags":16,"expectedKnowledge":26,"learningGoal":29,"keyTerms":30,"blocks":82},"it-aws-aws-certified-ai-practitioner-foundation-model-applications-foundation-model-training-fine-tuning-and-data","foundation-model-training-fine-tuning-and-data","基盤モデルの学習・微調整・データ準備","基盤モデルの学習・微調整・データ準備を、AIF-C01の公式objective 3.3.1・3.3.2・3.3.3に沿って、用語、判断順序、例、誤解修正で整理します。","\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Ffoundation-model-training-fine-tuning-and-data","\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications","beginner","cert_candidate",6,"2026-08-06","ai_generated",[17,18,19,20,21,22,23,24,25],"AWS Certified AI Practitioner","AIF-C01","基盤モデルの活用","事前学習","微調整","instruction tuning","transfer learning","RLHF","データキュレーション",[27,28],"AWSクラウドの基本用語を聞いたことがある","AIを使う業務例を日常的な言葉で考えられる","基盤モデルの学習・微調整・データ準備について、試験で問われる違いと選択基準を説明し、短い業務例へ適用できる。",[31,35,38,41,44,48,52,55,59,63,66,70,78],{"id":32,"label":21,"shortDefinition":33,"role":34},"fine-tuning","事前学習済みモデルを、追加データで特定用途へ適応させる学習です。","primary",{"id":36,"label":20,"shortDefinition":37,"role":34},"pretraining","広いデータから一般的なパターンを、利用前にモデルへ学習させる工程です。",{"id":39,"label":22,"shortDefinition":40,"role":34},"instruction-tuning","指示と望ましい応答の例で、指示追従の振る舞いを調整する学習です。",{"id":42,"label":23,"shortDefinition":43,"role":34},"transfer-learning","既に学んだ知識を、関連する別のタスクや領域へ活用する学習方法です。",{"id":45,"label":46,"shortDefinition":47,"role":34},"continuous-pretraining","継続事前学習","事前学習済みモデルへ追加の大規模データを学習させ、領域知識を深める方法です。",{"id":49,"label":50,"shortDefinition":51,"role":34},"model-distillation","モデル蒸留","高性能モデルの振る舞いを、より小さなモデルへ学習させる手法です。",{"id":53,"label":25,"shortDefinition":54,"role":34},"data-curation","目的に合うデータを収集・選別・修正し、品質を整える活動です。",{"id":56,"label":57,"shortDefinition":58,"role":34},"data-governance","データガバナンス","データの品質・権利・責任・アクセス・利用方法を統制する仕組みです。",{"id":60,"label":61,"shortDefinition":62,"role":34},"representativeness","代表性","対象となる利用者や状況の分布を、データが適切に反映している性質です。",{"id":64,"label":24,"shortDefinition":65,"role":34},"reinforcement-learning-from-human-feedback","人の好みや評価を信号として、モデルの振る舞いを調整する学習方法です。",{"id":67,"label":68,"shortDefinition":69,"role":34},"test-data","テストデータ","学習に使わず、未知ケースへの性能を評価するために分離したデータです。",{"id":71,"label":72,"shortDefinition":73,"supplement":74,"role":77},"foundation-model","基盤モデル","大規模なデータで事前学習され、多様な下流タスクへ適応できるモデルです。",{"kind":75,"text":76},"misconception","すべての定型予測を、基盤モデルへ置き換える必要はありません。","contextual",{"id":79,"label":80,"shortDefinition":81,"role":77},"evaluation","評価","定めたデータと基準で、モデルやアプリの品質・安全性・価値を測る活動です。",[83,88,100,109,118,124,130,136,152,161],{"id":84,"type":85,"title":7,"subtitle":86,"shortDefinition":87},"hero","HeroBlock","目的に合うデータ、調整方法、評価境界を設計する","事前学習済みモデルへ追加の例や指示データを使って学習を行い、特定タスクや領域での振る舞いを適応させる方法です。",{"id":89,"type":90,"term":91,"definition":92,"plainExplanation":93,"presentation":94,"supportLabel":95,"keywords":96},"definition","DefinitionBlock","微調整（fine-tuning）","[[cloze:fine-tuning|微調整]]とは、事前学習済みモデルを追加データで特定用途へ適応させる学習です。","[[term:foundation-model|基盤モデル]]は[[term:pretraining|事前学習]]で一般能力を得ます。[[term:instruction-tuning|instruction tuning]]は指示追従を、[[term:transfer-learning|transfer learning]]は既存知識の転用を狙います。[[term:continuous-pretraining|継続事前学習]]は領域知識を深め、[[term:model-distillation|モデル蒸留]]は小型モデルへ振る舞いを移します。","natural","言葉を分けると",[97,98,99],"追加データ","振る舞い適応","独立評価",{"id":101,"type":102,"steps":103,"title":108},"decision-steps","StepBlock",[104,105,106,107],"改善したい具体的な振る舞いと[[term:evaluation|評価]]基準を定義する。","権利と利用目的を確認し、[[term:data-curation|データキュレーション]]を行う。","[[term:data-governance|データガバナンス]]で品質と[[term:representativeness|代表性]]を点検し、[[term:test-data|テストデータ]]を学習から分ける。","調整前後を同じ評価で比較し、安全性と汎用能力の劣化も確認する。","調整データを準備する順序",{"id":110,"type":111,"title":112,"points":113,"body":117},"exam-cues","KeyPointBlock","判断で押さえる境界",[114,115,116],"調整目的と評価を先に定義","代表性・権利・品質・偏りを点検","学習と独立評価を分離","データ量だけを増やしても、誤った例や一部利用者だけの表現が多ければ品質は下がります。[[term:reinforcement-learning-from-human-feedback|RLHF]]も、人の好みや基準差を含み得ます。評価データの混入と評価者の偏りを監査します。",{"id":119,"type":120,"src":121,"alt":122,"caption":123},"diagram","DiagramBlock","\u002Fassets\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Ffoundation-model-training-fine-tuning-and-data\u002Ffoundation-model-training-fine-tuning-and-data.svg","基盤モデルの学習・微調整・データ準備について目的・基準、データ選別、微調整、独立評価、監視・改善の関係を文字ラベルと矢印で示した図","調整データの準備から独立評価、運用監視を循環させる図です。評価データを学習へ混ぜない境界を意識します。",{"id":125,"type":126,"scenario":127,"explanation":128,"result":129},"worked-example","ExampleBlock","社内サポート文体へFMを微調整するため、過去メールを集めます。","顧客の個人情報を除き、利用権限を確認し、誤った回答や古い規程を排除します。製品・言語・問い合わせ難度が偏らないように選び、学習用と最終テストを分けます。担当者が修正した高品質例を使い、調整後も正確さ、安全性、一般質問への劣化を比較します。","文体改善だけを狙いながら、機密漏えいと誤知識の学習を抑えられます。",{"id":75,"type":131,"warningTitle":132,"message":133,"severity":134,"action":135},"WarningBlock","過去ログはそのまま正解集ではない","古い手順、担当者の誤り、個人情報、偏った利用者層が含まれます。収集済みであることは、学習利用の権利と品質を保証しません。","warning","各データについて出所、利用権限、正しさ、代表性、評価への混入有無を記録します。",{"id":137,"type":138,"title":139,"questions":140},"quiz","QuizBlock","AIF-C01 判断チェック",[141],{"question":142,"choices":143,"answerIndex":147,"choiceExplanations":148},"微調整の性能を正直に測る方法はどれですか。",[144,145,146],"学習に使っていない代表テストデータで調整前後を比較する","学習例だけを再度採点する","データ出所を記録せず量だけ増やす",0,[149,150,151],"正解です。未知ケースへの一般化を確認します。","記憶した例で過大評価しやすくなります。","権利・品質・偏りを管理できません。",{"id":153,"type":154,"title":155,"points":156},"summary","SummaryBlock","まとめ",[157,158,159,160],"事前学習と追加調整は目的が違う","微調整方法を改善したい振る舞いで選ぶ","データ品質・代表性・権利を管理","独立評価で調整前後と副作用を比較する",{"id":162,"type":163,"title":164,"nextTopics":165,"relatedKeywords":167},"next-action","NextActionBlock","次は「基盤モデルの評価方法と指標」へ",[166],"it-aws-aws-certified-ai-practitioner-foundation-model-applications-foundation-model-evaluation-methods-and-metrics",[168,169,170],"モデル評価","ROUGE","BLEU",{"path":10,"title":19,"description":172,"parentPath":173,"accentToken":174},"基盤モデルの選択、RAG、プロンプト、カスタマイズ、エージェント、評価をAIF-C01の設計判断として整理します。","\u002Fit\u002Faws\u002Faws-certified-ai-practitioner","cloud-amber",{"title":176,"description":177,"canonicalUrl":178,"ogImage":179},"基盤モデルの学習・微調整・データ準備 | AWS Certified AI Practitioner","AIF-C01対策として基盤モデルの学習・微調整・データ準備の定義、判断軸、具体例、誤解しやすい境界を公式試験範囲に沿って学びます。","https:\u002F\u002Fzeqnilo.com\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Ffoundation-model-training-fine-tuning-and-data","https:\u002F\u002Fzeqnilo.com\u002Fassets\u002Fsite\u002Fdefault-og.svg",[181,184,187,188,189],{"title":182,"path":183},"IT","\u002Fit",{"title":185,"path":186},"AWS","\u002Fit\u002Faws",{"title":17,"path":173},{"title":19,"path":10},{"title":7,"path":9},[],[192],{"id":166,"title":193,"summary":194,"canonicalPath":195,"categoryPath":10,"categoryTitle":19,"difficulty":11,"targetAudience":12,"estimatedMinutes":13,"publishedAt":14,"updatedAt":14,"reviewStatus":15,"tags":196},"基盤モデルの評価方法と指標","基盤モデルの評価方法と指標を、AIF-C01の公式objective 3.4.1・3.4.2に沿って、用語、判断順序、例、誤解修正で整理します。","\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Ffoundation-model-evaluation-methods-and-metrics",[17,18,19,168,169,170,197,198,199],"BERTScore","LLM-as-a-judge","Amazon Bedrock Model Evaluation",[201],{"id":202,"title":203,"summary":204,"canonicalPath":205,"categoryPath":10,"categoryTitle":19,"difficulty":11,"targetAudience":12,"estimatedMinutes":13,"publishedAt":14,"updatedAt":14,"reviewStatus":15,"tags":206},"it-aws-aws-certified-ai-practitioner-foundation-model-applications-prompt-quality-risks-and-version-management","プロンプト品質・リスク・版管理","プロンプト品質・リスク・版管理を、AIF-C01の公式objective 3.2.3・3.2.4・3.2.5に沿って、用語、判断順序、例、誤解修正で整理します。","\u002Fit\u002Faws\u002Faws-certified-ai-practitioner\u002Ffoundation-model-applications\u002Fprompt-quality-risks-and-version-management",[17,18,19,207,208,209,210,211],"プロンプト管理","prompt injection","jailbreak","Amazon Bedrock Prompt Management","版管理",1787370539889]