{ "format": "agent-article-v0", "article_id": "EA-081", "canonical_url": "/tools/ea-081-fine-tuning-adapter-distillation", "page_type": "reference", "claim_ids": [], "evidence_ids": [ "EVD-EA-081-URL-CAE140A2C5E7", "EVD-EA-081-URL-2E95EE3FF8D6", "EVD-EA-081-URL-C31CB77E5FCF", "EVD-EA-081-URL-5AD14EFCBCF5", "EVD-EA-081-URL-5F95CF97560C", "EVD-EA-081-URL-D712A5AD2753", "EVD-EA-081-URL-59361017EB4C", "EVD-EA-081-URL-762EE5091783", "EA-081-R11-RESEARCH", "EA-081-R11-WEB-READBACK" ], "unknown_ids": [ "EA-081-UNKNOWN-01", "EA-081-UNKNOWN-02", "EA-081-UNKNOWN-CURRENTNESS" ], "operational_object_ids": [], "provider": [], "models": [], "plans": [], "surfaces": [], "regions": [], "current_status": "UNKNOWN", "observation_status": "SPEC_ONLY", "freshness": "UNKNOWN", "last_verified": null, "supersedes": [], "superseded_by": [], "provenance": [ "COMM-076-EA0081", "SRC-EA-081-URL-CAE140A2C5E7", "SRC-EA-081-URL-2E95EE3FF8D6", "SRC-EA-081-URL-C31CB77E5FCF", "SRC-EA-081-URL-5AD14EFCBCF5", "SRC-EA-081-URL-5F95CF97560C", "SRC-EA-081-URL-D712A5AD2753", "SRC-EA-081-URL-59361017EB4C", "SRC-EA-081-URL-762EE5091783", "WEBBODY-EA0081-R8-AF60D5C6" ], "article_sources": [ { "id": "EVD-EA-081-URL-CAE140A2C5E7", "evidence_class": "RESEARCH", "title": "[E1] source_class=primary paper | title=LoRA: Low-Rank Adaptation of Large Language Models | author/org=Edward J. Hu et al. | canonical=", "href": "https://arxiv.org/abs/2106.09685", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-CAE140A2C5E7" ], "supporting_locators": [ "EVIDENCE TABLE", "E1", "[E1] source_class=primary paper | title=LoRA: Low-Rank Adaptation of Large Language Models | author/org=Edward J. Hu et al. | canonical=https://arxiv.org/abs/2106.09685" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-2E95EE3FF8D6", "evidence_class": "RESEARCH", "title": "[E2] source_class=primary paper | title=QLoRA: Efficient Finetuning of Quantized LLMs | author/org=Tim Dettmers et al. | canonical=", "href": "https://arxiv.org/abs/2305.14314", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-2E95EE3FF8D6" ], "supporting_locators": [ "EVIDENCE TABLE", "E2", "[E2] source_class=primary paper | title=QLoRA: Efficient Finetuning of Quantized LLMs | author/org=Tim Dettmers et al. | canonical=https://arxiv.org/abs/2305.14314" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-C31CB77E5FCF", "evidence_class": "RESEARCH", "title": "[E3] source_class=primary paper | title=Distilling the Knowledge in a Neural Network | author/org=Geoffrey Hinton, Oriol Vinyals, Jeff Dean | canonical=", "href": "https://arxiv.org/abs/1503.02531", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-C31CB77E5FCF" ], "supporting_locators": [ "EVIDENCE TABLE", "E3", "[E3] source_class=primary paper | title=Distilling the Knowledge in a Neural Network | author/org=Geoffrey Hinton, Oriol Vinyals, Jeff Dean | canonical=https://arxiv.org/abs/1503.02531" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-5AD14EFCBCF5", "evidence_class": "RESEARCH", "title": "[E4] source_class=author-maintained implementation/docs | title=PEFT documentation — LoRA | author/org=Hugging Face | canonical=", "href": "https://huggingface.co/docs/peft/main/en/conceptual_guides/lora", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-5AD14EFCBCF5" ], "supporting_locators": [ "EVIDENCE TABLE", "E4", "[E4] source_class=author-maintained implementation/docs | title=PEFT documentation — LoRA | author/org=Hugging Face | canonical=https://huggingface.co/docs/peft/main/en/conceptual_guides/lora" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-5F95CF97560C", "evidence_class": "RESEARCH", "title": "[E5] source_class=official product docs | title=Fine-tuning guide | author/org=OpenAI | canonical=", "href": "https://platform.openai.com/docs/guides/fine-tuning", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-5F95CF97560C" ], "supporting_locators": [ "EVIDENCE TABLE", "E5", "[E5] source_class=official product docs | title=Fine-tuning guide | author/org=OpenAI | canonical=https://platform.openai.com/docs/guides/fine-tuning" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-D712A5AD2753", "evidence_class": "RESEARCH", "title": "[E6] source_class=official product docs | title=Model Distillation | author/org=OpenAI | canonical=", "href": "https://platform.openai.com/docs/guides/distillation", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-D712A5AD2753" ], "supporting_locators": [ "EVIDENCE TABLE", "E6", "[E6] source_class=official product docs | title=Model Distillation | author/org=OpenAI | canonical=https://platform.openai.com/docs/guides/distillation" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-59361017EB4C", "evidence_class": "RESEARCH", "title": "[E7] source_class=official product docs | title=Tune Gemini models by using supervised fine-tuning | author/org=Google Cloud | canonical=", "href": "https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-59361017EB4C" ], "supporting_locators": [ "EVIDENCE TABLE", "E7", "[E7] source_class=official product docs | title=Tune Gemini models by using supervised fine-tuning | author/org=Google Cloud | canonical=https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-081-URL-762EE5091783", "evidence_class": "RESEARCH", "title": "[E8] source_class=official product docs | title=Customize your model to improve its performance for your use case | author/org=Amazon Web Services | canonical=", "href": "https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-081-URL-762EE5091783" ], "supporting_locators": [ "EVIDENCE TABLE", "E8", "[E8] source_class=official product docs | title=Customize your model to improve its performance for your use case | author/org=Amazon Web Services | canonical=https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EA-081-R11-RESEARCH", "evidence_class": "RESEARCH", "title": "EA-081 監査済みresearch / canonical revision binding", "supports": [ "F1. “Fine-tuning”を一語で扱わない。最低でも (a) full/managed supervised fine-tuning、(b) adapter/LoRA PEFT、(c) quantized PEFT/QLoRA、(d) distillation を別artifact classとする。E1-E3。\nF2. LoRA/adapterはbase modelとadapter artifactの組合せとして扱う必要がある。base model revisionが変わればadapterの意味・互換性・性能は再検証対象になる。これはE1の構造から導くIMPLEMENTATION INFERENCEであり、全製品への互換保証/非保証を断定しない。\nF3. QLoRAは“LoRAの安価版”というだけではない。quantization configuration、base weights、adapter、training stackが再現性contractに入る。E2。\nF4. Distillationはteacher output/behaviorをstudent trainingへ利用する関係であり、studentがteacherの能力、安全性、知識鮮度、tool behaviorを完全継承することを意味しない。E3/E6。\nF5. managed customizationはprovider-level capabilityではなく、model ID × method × region × account/project permission × quota × API/service revisionのavailability gateで評価する。E5-E8。\nF6. prompt/RAG/tuningは代替関係だけではない。頻繁に変わる知識・ACL付き企業情報は外部retrieval側、安定した出力様式・task behaviorはtuning候補になり得るが、採用は同一holdoutでbaseline比較して決める。PROPOSAL。\nF7. training lossの改善はdeployment acceptanceではない。untuned baseline、prompt/RAG baseline、tuned artifactを同じhidden holdoutで比較し、critical regressionsを平均scoreで相殺しない。" ], "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。", "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET", "source_or_experiment_ids": [ "COMM-076-EA0081" ], "supporting_locators": [ "COMM-076-EA0081", "EA-081-R5-LEGACY-MIGRATION-20261008-R1", "content/staging/EA-081.md" ], "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。", "provenance_status": "AUDIT_PASS_CANONICAL_BINDING" }, { "id": "EA-081-R11-WEB-READBACK", "evidence_class": "IMPL", "title": "EA-081 Web body identity / canonical readback", "supports": [ "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。" ], "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。", "checked_at": null, "origin_class": "IMPLEMENTATION_READBACK", "source_or_experiment_ids": [ "WEBBODY-EA0081-R8-AF60D5C6" ], "supporting_locators": [ "WEBBODY-EA0081-R8-AF60D5C6", "content/staging/EA-081.md", "EA-081-R5-LEGACY-MIGRATION-20261008-R1" ], "scope": "R9/R10 canonical readback後のProduction生成時body identity。", "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH" } ], "research_sections": [ { "heading": "WHY_THIS_MATTERS_TO_UPDATE", "text": "企業AIでは「基盤モデルをそのまま使う」以外に、教師付きfine-tuning、LoRA等のparameter-efficient fine-tuning (PEFT)、量子化併用PEFT、knowledge distillationなど複数の適応経路が実用化されている。これらは同じ“カスタマイズ”ではなく、更新対象パラメータ、学習資源、serving artifact、teacher依存、再学習条件、rollback可能性、データ境界が異なる。公開後版ではprompt/RAGとの選択境界、provider managed customizationとOSS/self-hosted customizationの差、評価・監査・コストを独立して扱う必要がある。" }, { "heading": "SCOPE_ROUTING", "text": "Enterprise AI Core / provider mapping / organization overlay / implementation / governance / benchmark-evaluation / security-privacy / cost-operability" }, { "heading": "CLAIM_OR_RESEARCH_QUESTION", "text": "1. Fine-tuning、adapter/LoRA/QLoRA、distillationをどのartifact/lifecycle境界で分離すべきか。\n2. “学習できた”を“本番で安全・有効”へ外挿しないため、何を固定・比較・監査すべきか。\n3. provider managed tuningの存在と、特定model/region/plan/APIでのavailabilityをどう分離するか。\n4. teacherの品質をstudentが自動的に継承するという誤解をどう防ぐか。" }, { "heading": "EVIDENCE TABLE", "text": "[E1] source_class=primary paper | title=LoRA: Low-Rank Adaptation of Large Language Models | author/org=Edward J. Hu et al. | canonical=https://arxiv.org/abs/2106.09685 | published=2021-06-17 | accessed=2026-09-16 | version=arXiv:2106.09685 | locator=Abstract/Method | label=RESEARCH | relation=support/context | establishes freezing pretrained weights while injecting trainable low-rank matrices; parameter-efficient adaptation is structurally distinct from full fine-tuning.\n[E2] source_class=primary paper | title=QLoRA: Efficient Finetuning of Quantized LLMs | author/org=Tim Dettmers et al. | canonical=https://arxiv.org/abs/2305.14314 | published=2023-05-23 | accessed=2026-09-16 | version=arXiv:2305.14314 | locator=Abstract/Methods | label=RESEARCH | relation=support/limit | establishes backpropagation through a frozen quantized pretrained model into LoRA adapters; quantization+adapter is a distinct implementation path, not equivalent to full-precision full-model tuning.\n[E3] source_class=primary paper | title=Distilling the Knowledge in a Neural Network | author/org=Geoffrey Hinton, Oriol Vinyals, Jeff Dean | canonical=https://arxiv.org/abs/1503.02531 | published=2015-03-09 | accessed=2026-09-16 | version=arXiv:1503.02531 | locator=Abstract/Distillation section | label=RESEARCH | relation=support/context | establishes teacher/ensemble knowledge transfer to a smaller deployable model through training targets; distillation is a training relationship, not model identity equivalence.\n[E4] source_class=author-maintained implementation/docs | title=PEFT documentation — LoRA | author/org=Hugging Face | canonical=https://huggingface.co/docs/peft/main/en/conceptual_guides/lora | accessed=2026-09-16 | version=current web docs; exact package release not pinned in this run | locator=LoRA conceptual guide | label=OSS_PRACTICE / PRODUCT_VOLATILE | relation=support/context | current PEFT implementation exposes LoRA as adapter-based parameter-efficient tuning; exact package/version must be pinned by update Work before reproducibility claims.\n[E5] source_class=official product docs | title=Fine-tuning guide | author/org=OpenAI | canonical=https://platform.openai.com/docs/guides/fine-tuning | accessed=2026-09-16 | version=current docs; supported model IDs are PRODUCT_VOLATILE | locator=Fine-tuning methods/workflow/evaluation guidance | label=VERIFIED_SPEC / PRODUCT_VOLATILE | relation=support/limit | managed fine-tuning exists as provider workflow; current supported models, method availability and account eligibility must be re-read at adoption time.\n[E6] source_class=official product docs | title=Model Distillation | author/org=OpenAI | canonical=https://platform.openai.com/docs/guides/distillation | accessed=2026-09-16 | version=current docs; model IDs PRODUCT_VOLATILE | locator=Distillation workflow | label=VERIFIED_SPEC / PRODUCT_VOLATILE | relation=support/context | provider workflow can use stronger-model outputs/evals as part of a smaller-model optimization loop; this does not establish teacher-equivalent quality across arbitrary workloads.\n[E7] source_class=official product docs | title=Tune Gemini models by using supervised fine-tuning | author/org=Google Cloud | canonical=https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-models | accessed=2026-09-16 | version=current Vertex AI docs; model/region support PRODUCT_VOLATILE | locator=Supported models / tuning workflow | label=VERIFIED_SPEC / PRODUCT_VOLATILE | relation=support/limit | managed supervised tuning is provider/model-gated; availability must be evaluated by model, region, project permissions and service state rather than provider-level yes/no.\n[E8] source_class=official product docs | title=Customize your model to improve its performance for your use case | author/org=Amazon Web Services | canonical=https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html | accessed=2026-09-16 | version=current Bedrock docs; model/region support PRODUCT_VOLATILE | locator=Model customization methods | label=VERIFIED_SPEC / PRODUCT_VOLATILE | relation=support/limit | Bedrock exposes managed model customization paths, but supported customization method varies by model/provider/region and must be pinned at update time." }, { "heading": "FINDINGS", "text": "F1. “Fine-tuning”を一語で扱わない。最低でも (a) full/managed supervised fine-tuning、(b) adapter/LoRA PEFT、(c) quantized PEFT/QLoRA、(d) distillation を別artifact classとする。E1-E3。\nF2. LoRA/adapterはbase modelとadapter artifactの組合せとして扱う必要がある。base model revisionが変わればadapterの意味・互換性・性能は再検証対象になる。これはE1の構造から導くIMPLEMENTATION INFERENCEであり、全製品への互換保証/非保証を断定しない。\nF3. QLoRAは“LoRAの安価版”というだけではない。quantization configuration、base weights、adapter、training stackが再現性contractに入る。E2。\nF4. Distillationはteacher output/behaviorをstudent trainingへ利用する関係であり、studentがteacherの能力、安全性、知識鮮度、tool behaviorを完全継承することを意味しない。E3/E6。\nF5. managed customizationはprovider-level capabilityではなく、model ID × method × region × account/project permission × quota × API/service revisionのavailability gateで評価する。E5-E8。\nF6. prompt/RAG/tuningは代替関係だけではない。頻繁に変わる知識・ACL付き企業情報は外部retrieval側、安定した出力様式・task behaviorはtuning候補になり得るが、採用は同一holdoutでbaseline比較して決める。PROPOSAL。\nF7. training lossの改善はdeployment acceptanceではない。untuned baseline、prompt/RAG baseline、tuned artifactを同じhidden holdoutで比較し、critical regressionsを平均scoreで相殺しない。" }, { "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE", "text": "- catastrophic forgetting / capability regression: 狭いtaskへの適応で一般能力・安全挙動が落ち得る。特定手法の必然的failure rateは本Packetでは確定しない。\n- overfitting / leakage: train/eval重複、near-duplicate、template leakageで見かけ上の改善が生じる。\n- adapter/base drift: base revision変更後に旧adapter評価をそのまま継承しない。\n- distillation ceiling/shift: teacher error/biasをstudentへ転写し得る一方、studentの容量・training objectiveにより別のfailureも生じる。teacher pass ≠ student pass。\n- safety regression: task quality改善とpolicy/safety behaviorは別評価軸。\n- data governance: tuning corpusへの秘密情報・PII投入可否、retention、training data handlingはprovider/tenant契約・設定依存。public docsだけで特定tenant状態を証明しない。\n- benchmark transfer: public benchmark improvementをtenant workloadの保証へ外挿しない。" }, { "heading": "AVAILABILITY_GATES", "text": "provider; base model exact ID/revision; customization method; region; account/project/org eligibility; API/service revision; quota; minimum/maximum dataset constraints; supported input/output modality; adapter serving support; dedicated/shared endpoint requirement; encryption/data residency controls; retention; pricing; deprecation state. 2026-09-16時点のcurrent product matrixはPRODUCT_VOLATILEであり、公開後Workで再pinする。" }, { "heading": "IMPLEMENTATION_STATE", "text": "- Research/OSS: LoRAとQLoRAは確立した公開手法で、PEFT等に実装経路がある。RESEARCH / OSS_PRACTICE。\n- Managed provider: OpenAI/Vertex AI/Bedrockにmodel customization経路が存在する。VERIFIED_SPEC / PRODUCT_VOLATILE。ただし“全modelでfine-tuning可能”とはしない。\n- Enterprise Core proposal: customization artifactを base_model_ref + method + training_data_manifest + hyperparameters + code/service revision + resulting_artifact_id/hash + eval_suite + approval として登録する。" }, { "heading": "BENCHMARK_OR_EVAL", "text": "後続TEST_PROFILE:\nA. fixed hidden holdoutを、untuned base / prompt-only / RAG where applicable / tuned or adapter / distilled studentで比較。\nB. task quality: exact/task success, schema adherence, domain terminology, calibration where meaningful。\nC. regression: general capability probes, refusal/safety, tool/function-call correctness, multilingual, long-context subset。\nD. memorization/privacy: canary/near-verbatim extraction、train-neighbor test、PII/secret scan。\nE. robustness: paraphrase、OOD、adversarial/ambiguous inputs。\nF. operability: training wall time、training cost、artifact storage、serving latency p50/p95、per-request cost、rollback time。\nG. distillation: teacher vs student disagreement matrixを保持し、teacher-correct/student-wrong、teacher-wrong/student-correct、both-wrongを分離。\nAcceptanceは平均scoreだけでなくcritical-regression gateを持つ。" }, { "heading": "SECURITY/PRIVACY", "text": "training corpusのorigin、classification、consent/licensing、PII/secrets、provider upload path、retention/deletion evidenceを保持。公開資料をtenant実設定の証明にしない。\nAUDIT: dataset manifest/hash、split seed、base model ID、method、hyperparameters、service/package version、artifact ID/hash、eval suite/version、approval、deployment endpoint/version、rollback targetを記録。" }, { "heading": "COST", "text": "training費用だけでなく、data curation、eval、artifact storage、dedicated endpoint、inference token/compute、再学習頻度をTCO化。LoRA/QLoRAの研究上の効率性を特定managed productの料金優位へ外挿しない。" }, { "heading": "OPERABILITY", "text": "base retirement、adapter compatibility、provider model deprecation、dataset refresh、rollback、multi-region availabilityをlifecycleに含める。" }, { "heading": "SAFE_WORDING_FOR_FUTURE_BOOK", "text": "「モデル適応には、教師付きfine-tuning、LoRA等のparameter-efficient tuning、量子化併用PEFT、distillationなど異なる経路がある。これらは学習・配備artifactとfailure surfaceが異なるため、“カスタマイズ済み”を一括して扱わない。採用時はbase model、method、training data、artifact、evaluation、availabilityをversion付きで固定し、未調整baselineと同一holdoutで品質・安全性・privacy・costを再評価する。Distillationはteacherとstudentの同一性を保証しない。」" }, { "heading": "UPDATE_WORK_HANDOFF", "text": "1. 公開後更新時にOpenAI / Google Vertex AI / AWS Bedrockのcurrent supported model × customization method × region × pricing matrixを公式docsから再pin。\n2. OSS採用候補ではPEFT/Transformers/bitsandbytes等のexact release/tag/commitと互換matrixを固定。\n3. tuning corpusのgovernance contractとeval split contamination checkを先に作る。\n4. `MODEL CUSTOMIZATION ARTIFACT CONTRACT` をCore候補として検討: base_model_ref, tokenizer/processor_ref, method, adapter/quantization config, dataset manifest/hash, training code/service revision, hyperparameters, resulting artifact ID/hash, eval suite, safety/privacy results, approval, deployment/rollback refs。\n5. prompt/RAG/tuning/distillationを同一task economicsで比較し、training costだけで選ばない。" }, { "heading": "OPEN_UNKNOWN", "text": "/ STOP_CONDITION" }, { "heading": "OPEN_UNKNOWN", "text": "2026-09-16時点の各providerのexact supported model IDs、region、pricing、quota、retention/data-residency matrix; PEFT current exact release/tag/commit; managed provider内部がLoRA等どのparameterizationを採るか（providerが明示しない限りUNKNOWN）。" }, { "heading": "STOP_CONDITION", "text": "本Packetでは手法境界、主要一次研究、managed provider existence、failure envelope、evaluation/audit contractまで確立。次回同じEA-081を再採集するのは、上記OPEN_UNKNOWNをcurrent provider mappingとして閉じる場合、またはsource version/claim relationに実質差分が生じた場合のみ。\n\n---" }, { "heading": "AUDITED_CURRENTNESS_BOUNDARY", "text": "OpenAI fine-tuning is a shrinking PRODUCT_VOLATILE path; provider-managed fine-tuning must not be represented as stable future default or as exposing LoRA/adapter internals." } ], "human_url": "/tools/ea-081-fine-tuning-adapter-distillation", "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-081-fine-tuning-adapter-distillation", "json_url": "/agent/article.json?path=%2Ftools%2Fea-081-fine-tuning-adapter-distillation" }