2nd/10_Wiki/Topics/AI_and_ML/SFT (Supervised Fine-Tuning).md

id, title, category, status, canonical_id, duplicate_of, aliases, source_trust_level, confidence_score, verification_status, tags, last_reinforced, github_commit

id	title	category	status	canonical_id	duplicate_of	aliases	source_trust_level	confidence_score	verification_status	tags	last_reinforced	github_commit
wiki-2026-0508-sft-supervised-fine-tuning	SFT (Supervised Fine-Tuning)	10_Wiki/Topics	duplicate	fine-tuning	Fine-tuning	Supervised Fine-Tuning Instruction Tuning	A	0.9	redirected	duplicate sft fine-tuning llm	2026-05-10	pending

SFT (Supervised Fine-Tuning)

이 문서는 Fine-tuning 의 중복본입니다. Canonical 문서로 redirect.

핵심 요약 (SFT-specific aspects)

• SFT = supervised stage of LLM post-training (prompt → response pairs).
• 매 RLHF/DPO pipeline 의 첫 stage — base model → instruction-following model.
• 일반적 dataset: ShareGPT, OpenAssistant, Alpaca-style, custom domain Q&A.
• 매 2026 typical recipe: LoRA/QLoRA on Llama 3.x / Qwen 2.5 + 1-3 epochs at ~2e-5 LR.
• 후속 stage: DPO → KTO → online RL (GRPO).

🔗 Graph

• 부모: Fine-tuning (canonical)
• Adjacent: RLHF · DPO · LoRA · Instruction-Tuning

🕓 변경 이력

날짜	변경
2026-05-08	Phase 1
2026-05-10	중복 처리 — canonical 문서로 redirect