AI Post Training

AI Post-Training

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 (ai-architecture-advisor), or serving the result (ai-llm-inference).

Quick Reference

When to Use This Skill

Activate when the user asks (in any language) some form of:

• "How do I run RLHF / align a model / train with human feedback?"
• "DPO vs PPO vs GRPO — which preference/RL method?"
• "How do I train a reasoning model / RLVR / GRPO like DeepSeek-R1?"
• "How do I build/choose a reward model? ORM or PRM?"
• "Should I use Constitutional AI / RLAIF instead of human labels?"
• "My fine-tune still has a preference/safety/refusal gap after SFT — now what?"
• "How do I collect preference data / what about synthetic preference data?"
• "My RL model is reward-hacking / over-refusing — how do I fix over-optimization?"

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.

Scope Boundaries (Use These Skills for Depth)

• Per-algorithm catalogue + decision tree (PPO/DPO/GRPO/RLVR/RULER/...) -> ai-llm/references/post-training.md
• TRL / SFT / DPO / GRPO implementation in code -> huggingface-skills: plugin (TRL)
• Distributed RL training scale (FSDP, vLLM rollout, async RL) -> ai-distributed-training
• Eval methodology, judge calibration, thresholds -> ai-evals
• The prompt→RAG→SFT→post-train promotion decision -> ai-architecture-advisor
• Reasoning-model build walkthrough -> Raschka, Build a Reasoning Model (see sources)

Workflow

1. Confirm post-training is the right rung. Is the gap knowledge (→ RAG), format/behavior demonstrable with labels (→ SFT), or reasoning closeable on a hosted model (→ raise the thinking budget)? If yes to any, stop — you don't need post-training. → verify: name the gap type.
2. Exhaust SFT. Establish the SFT baseline; only proceed if a measurable preference/safety/ reasoning gap remains. → verify: SFT eval shows the residual gap.
3. Identify the reward signal you can actually produce — human pairs, AI feedback, a written rubric, or a verifiable checker. This, not a benchmark, picks the algorithm. → verify: signal is real and labelable.
4. Pick the method (see Choosing the Method): on-policy distillation if a stronger teacher exists; otherwise offline DPO/DAAs first, promote to GRPO/RLOO online on evidence, RLVR when the reward is verifiable. → verify: simplest method that fits the signal.
5. Build/choose the reward model or checker (see Reward Modeling). → verify: RM accuracy or checker coverage.
6. Train with an eval harness from step 1, and KL scoped to the reward source (KL when the reward is learned; β=0 with a verifiable checker). → verify: held-out true-objective metric, not reward curve.
7. Hand off per-algorithm depth to ai-llm/references/post-training.md and scale to ai-distributed-training.

The Post-Training Pipeline

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 + GRPO/RLOO (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:

• Online vs offline. Offline (DPO/DAAs) trains on a fixed preference dataset — simple, stable, no sampling loop or reward model. Online (GRPO/RLOO, or historically PPO) samples from the current policy and scores it live — higher ceiling, more compute and moving parts. Start offline; go online when offline plateaus or you need a reward model's generalization.
• Reward source. Human preferences → reward model; AI preferences → RLAIF/Constitutional AI; a written multi-criteria rubric → rubrics-as-rewards; verifiable checker (compiler, unit tests, math solver) → RLVR. The reward source you can actually produce determines the algorithm more than any benchmark does — and it also determines whether KL is your trust region (learned reward) or clipping is (verifiable checker).
• Reference-based vs reference-free. Within the DAA family, DPO/KTO keep a frozen reference model (memory cost, implicit drift bound); ORPO/SimPO drop it (cheaper, no drift bound — pair with a capability regression suite).

Choosing the Method

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:

1. Can you write demonstrations? → SFT first. Do not reach for RL to teach something a few hundred labeled examples would teach.
2. Do you have pairwise preferences and want simplicity? → DPO (then KTO/ORPO/SimPO if its numerics misbehave or you only have binary good/bad signals).
3. Does a stronger teacher model already exist, with a small student? → on-policy distillation before any RL loop: the teacher scores the student's own rollouts token-by-token (on-policy, dense). Reported to outperform SFT and GRPO in that setting and to restore generalization SFT loses.
4. Can you afford a reward model + online RL for a higher ceiling? → a critic-free group-baseline method (GRPO/RLOO) is the 2026 default; it drops the value model and its optimizer state. PPO remains the reference algorithm (InstructGPT lineage) but ships under trl.experimental — a learned reward model does not imply PPO.
5. Is the reward verifiable (math/code/tests)? → RLVR, usually via GRPO or a GRPO-family variant (DAPO/GSPO/RLOO) — the dominant 2026 reasoning recipe, now a portfolio rather than one fixed algorithm; no human labels needed.
6. Is the task real work with no mechanical checker? → rubrics as rewards: a structured multi-criteria rubric grades the response. Legible and auditable, but a model-mediated proxy — so the KL and over-optimization controls apply as they do for a reward model.
7. Are human labels the bottleneck? → RLAIF / Constitutional AI to generate the preference/critique signal from a model + a written constitution.
8. Want a quick gain without an RL loop? → Rejection sampling: best-of-N generate → score → SFT on the winners.

Reward Modeling (the load-bearing component)

In reward-model-based RLHF, model quality is capped by reward-model quality. Key choices:

• Bradley-Terry RM — the standard: an LM with a scalar value head trained on preference pairs to predict which response a human prefers. Quality depends on preference-data balance and avoiding spurious length/format correlations.
• ORM vs PRM — Outcome Reward Models score the final answer; Process Reward Models score each reasoning step. PRMs help on multi-step reasoning but need step-level labels and are costlier to build. PRMs themselves split into discriminative (a scalar per step — the 2023 form, brittle on step segmentation and documented as hackable) and generative (the verifier reasons, then judges — the 2026 default where PRMs are used at all).
• Generative reward modeling / LLM-as-a-judge — use a model to emit a critique or score instead of a scalar head; flexible, but inherits the judge's biases (calibrate via ai-evals).
• For RLVR you skip the reward model — a deterministic checker is the reward. That is why RLVR is cheaper and harder to over-optimize than reward-model RL where the checker exists. The checker is a much tighter proxy, not the true objective: incomplete tests are still hackable.
• Rubrics as rewards — when the task is real work with no mechanical checker, a structured multi-criteria rubric can be the reward instead of forcing a fake verifier or falling back to opaque pairwise preferences. Still a model-mediated proxy; treat it like a reward model for over-optimization purposes.

Depth: references/reward-and-data.md.

Over-Optimization Is the Default Failure Mode

Preference RL optimizes a proxy for what you want, so it Goodharts silently — the model games the reward while the true objective degrades. Controls:

• KL regularization — scoped by reward source. With a learned reward (RM+PPO, rubric grader, DPO's implicit β) KL to the reference policy is the primary trust region and the main knob against reward hacking: tune it, don't omit it. Under a verifiable checker (RLVR), beta=0 is the 2026 standard — TRL's GRPOConfig ships beta=0.0, DAPO drops the KL term, GSPO sets it to zero — and the trust region is carried by PPO-style clipping instead. Reach for a nonzero β there only on evidence of drift or capability regression.
• Eval harness, always — "completed" is wrong if anything was skipped; measure the true objective (held-out human eval / verifiable tests), not just rising reward. Watch for over-refusal (the model refuses safe requests) and length/sycophancy inflation.
• On-policy data + pretraining-gradient mixing — mitigate forgetting and distribution collapse.

Depth: references/over-optimization-and-eval.md.

Known Traps

• reaching for PPO/GRPO when DPO would do — paying for a reward model + RL loop you don't need
• post-training at all when the gap is missing knowledge (RAG) or format (SFT), not preference/reasoning
• treating RLHF as one algorithm — it's a pipeline (SFT → preference → reasoning RL) with online/offline and reward-source choices inside it
• training a reward model on imbalanced/length-correlated preferences, then optimizing its spurious signal
• running preference RL without an eval harness — reward goes up, true quality goes down, silently (Goodhart)
• omitting the KL penalty in reward-model RL and watching the policy drift off its trusted SFT behavior (reward hacking, over-refusal) — but carrying a nonzero KL into RLVR by reflex, where β=0 is standard and KL mostly caps the reasoning gain
• carrying a beta value across method families — DPO's β (~0.1, an implicit-reward temperature) and a GRPO KL coefficient (0.0–0.001) are different objects two orders of magnitude apart
• reaching for GRPO when a stronger teacher already exists — on-policy distillation is the cheaper and often better move for a small student
• picking among DPO/KTO/ORPO/SimPO from a list of adjectives instead of the reference-based vs reference-free tradeoff (a frozen model in memory and an implicit drift bound, or neither)
• assuming a single-turn RLVR recipe transfers to a multi-turn agent — trajectory-level reward, cross-turn credit assignment, and rollout infrastructure are all new problems
• using RLVR where the reward is not actually verifiable (no deterministic checker) — then it's just reward-model RL with a brittle checker
• confusing ORM and PRM — process rewards need step-level labels you may not have
• running vanilla GRPO on a large MoE and fighting non-convergence — token-level ratios break under expert-routing volatility; use GSPO (sequence-level)
• ignoring GRPO's length/std biases that inflate response length and miscalibrate difficulty — use Dr. GRPO / DAPO fixes (see methods reference)
• assuming a reasoning gap needs RLVR when, on a hosted model, raising the thinking budget would close it without any training

Common Anti-Patterns

• jumping to RL before SFT is exhausted
• choosing the algorithm from a benchmark instead of from the reward signal you can produce
• treating reward-model quality as an afterthought when it caps the whole result
• measuring success by reward curve instead of the true held-out objective
• this skill re-teaching the per-algorithm math instead of routing to the ai-llm catalogue

Core Principles

1. SFT first, RL last. Exhaust demonstrations before any reward-based method.
2. The reward signal picks the algorithm. Four sources: human pairs → DPO/RM+GRPO; AI preferences → RLAIF; a rubric → rubrics-as-rewards; a verifiable checker → RLVR.
3. Offline before online. Start with DPO's simplicity; promote to GRPO/RLOO on evidence.
4. Reward quality caps model quality. Invest in the reward model, rubric, or checker accordingly.
5. Assume over-optimization. Always eval the true objective, or it Goodharts. Add KL to the reference when the reward is learned; under a verifiable checker the trust region is clipping and β=0 is standard.

Navigation: Core References

• methods-and-pipeline.md — the SFT→preference→RL pipeline, online vs offline, reference-based vs reference-free, and how each method (DPO/PPO/GRPO/RLVR/rejection sampling/on-policy distillation) maps to a reward signal; also agentic/multi-turn RL, rubrics-as-rewards, and the per-method beta anchor table; routes to the ai-llm algorithm catalogue for per-algorithm depth
• reward-and-data.md — reward modeling (Bradley-Terry, ORM/PRM, generative RM), preference-data collection, synthetic data, RLAIF/Constitutional AI
• over-optimization-and-eval.md — reward hacking/Goodhart, KL regularization, over-refusal, and evaluating the true objective
• grpo-run-diagnostics.md — reading a live GRPO/RLVR run: advantage mean (sanity check) vs std (learning signal), degenerate zero-gradient groups, reward exhaustion at 1.00, entropy trajectories, and a triage table

External Sources

See data/sources.json for primary references: Lambert's RLHF book (the anchor), InstructGPT, DPO, DeepSeek-R1 (GRPO/RLVR), Tülu 3, GKD and Thinking Machines' on-policy distillation, Rubrics as Rewards, the multi-turn agentic RL practitioner's guide, the PRM survey, Raschka's Build a Reasoning Model (verifier engineering + GRPO run telemetry), and Pai's Designing Large Language Model Applications (model merging/fusion taxonomy).

Fact-Checking

• Algorithm names, framework support, and which labs use which recipe are volatile; verify against current primary sources before recommending a specific one. TRL specifically turns over fast — its loss_type roster, trainer namespaces (first-class vs trl.experimental), and defaults all changed between 2026-07 and 2026-08.
• The framework landscape is wider than TRL: verl (the common backbone for large-scale and agentic RL, async rollout), OpenRLHF (multi-turn/VLM RL), and others (NeMo RL, AReaL, ROLL, slime). Choose beyond TRL when scale, asynchronous rollout, or multi-turn environments are the constraint; delegate depth to ai-distributed-training. Health and feature claims for any of these must be re-checked — they were not verified past 2026-08.
• Model-specific recipe claims (e.g. "DeepSeek-R1 used X") must be checked against the model's own technical report, not secondary summaries.
• If you cannot verify, present guidance as a dated assumption, not a fact.

Learnings Loop

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.