메뉴
Post-training and alignment: reward modeling, RLHF/PPO, DPO/DAAs, GRPO, RLVR, RLAIF, over-optimization. Use when adapting an SFT model with preference or verifiable-reward signals.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/vasilyu1983/AI-Agents-public --skill ai-post-training명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | ai-post-training |
| description | Post-training and alignment: reward modeling, RLHF/PPO, DPO/DAAs, GRPO, RLVR, RLAIF, over-optimization. Use when adapting an SFT model with preference or verifiable-reward signals. |
| compatibility | Portable core. Works on Claude Code and Codex. |
| version | 1.1 |
| last_validated | 2026-07-11T00:00:00.000Z |
Domain: the rung after supervised fine-tuning — turning a pretrained or SFT'd base model into an aligned, preference-tuned, or reasoning-capable model with a reward signal. This skill owns the post-training decision and pipeline: when to post-train at all, which reward signal you can produce, which algorithm family fits, and how to keep it from over-optimizing. Per-algorithm operational depth lives in ai-llm/references/post-training.md (PPO, DPO, SimPO, KTO, GRPO, GSPO, DAPO, RLVR, RULER, ORPO — catalogue + decision tree); this skill routes there.
It does not cover: pretraining (ai-pretraining), the prompt→RAG→SFT promotion ladder (), or serving the result ().
| You have / want | Method | Deep ref |
|---|---|---|
| Labeled demonstrations of the target behavior | SFT (baseline — exhaust it first; not RL) | ai-llm |
| Pairwise preferences, want the least machinery | DPO (or DAAs: KTO / ORPO / SimPO) | methods |
| Preferences + reward model + online RL, highest ceiling | PPO (reward model + policy + critic) | methods |
| Many samples scorable per prompt, drop the critic | GRPO (group-relative advantage) | methods |
| A verifiable checker (math/code/tests) as the reward | RLVR (via GRPO or a GRPO-family variant — GSPO/DAPO/RLOO) — the dominant 2026 reasoning recipe | methods |
| Scale preference labels cheaply | RLAIF / Constitutional AI (model-as-judge) | data |
| A quick lift with no RL loop | Rejection sampling (best-of-N → SFT) | methods |
| Train/choose the reward model itself | Bradley-Terry RM, ORM vs PRM, generative RM | reward |
| Stop reward hacking / over-refusal | KL regularization, eval harness, over-optimization controls | over-optimization |
Activate when the user asks (in any language) some form of:
If the gap is missing knowledge (→ RAG), missing format/behavior demonstrable with labels (→ SFT), or reasoning closeable by more thinking on a hosted model (→ raise the thinking budget), you usually do not need this skill. Confirm with ai-architecture-advisor first if unsure.
huggingface-skills: plugin (TRL)Post-training is a sequence, not a single algorithm. Each stage is reached only when the previous one is exhausted and a measurable gap remains.
pretrained base
|
v
1. SFT (instruction tuning) teach the format/behavior from demonstrations
| gap remains: preferences, safety, style the labels can't express
v
2. preference optimization DPO / DAAs (offline) OR reward model + PPO/GRPO (online)
| gap remains: multi-step reasoning, verifiable correctness
v
3. reasoning RL (RLVR) verifiable rewards (math/code/tests), usually via GRPO
|
v
aligned / reasoning model + continuous eval against over-optimization
Two orthogonal choices run through stages 2–3:
Pick by the reward signal you can produce, then by compute budget. Full per-algorithm detail and a decision tree are in ai-llm/references/post-training.md; the front-door logic:
In reward-model-based RLHF, model quality is capped by reward-model quality. Key choices:
Depth: references/reward-and-data.md.
Preference RL optimizes a proxy for what you want, so it Goodharts silently — the model games the reward while the true objective degrades. Controls:
See data/sources.json for primary references: Lambert's RLHF book (the anchor), InstructGPT, DPO, DeepSeek-R1 (GRPO/RLVR), Tülu 3, and Raschka's Build a Reasoning Model.
Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).
After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.