Reference · vocabulary

From the book to the platform

Nathan Lambert's RLHF book, chapter by chapter, against what is here. The book's word, the platform's word, the call. Where the platform does not have the thing, the row says so.

Chapter by chapter

ChapterThe book saysHereCall or pageOn this site
01 IntroductionRLHFmeasure, simulate, train, prove-Workflow
02 A Tiny History of RLHFhistory of RLHFnot here-Post-training
03 Training OverviewSFT, reward model, RL stagesTrain button: SFT, GRPO, DPOzps.train(dataset, method="sft" | "grpo" | "dpo")Training
04 Instruction Fine-Tuninginstruction tuning, loss maskingSFT rows with a loss maskzps.export_training(rows, mask_mode="assistant" | "final")Training rows
05 Reward Modelingreward modeltrained on the pairs, then a judgezps.train(ds, method="rm"); zps.reward_model(run)Reward model
06 Reinforcement Learningpolicy gradient: PPO, GRPOGRPO only, no PPOzps.train(dataset, method="grpo")RL, end to end
07 Reasoning and Inference-Time ScalingRLVR, pass@kverifiable reward, pass@k with intervalszps.pass_at(rows, k=8)Three numbers
08 Direct-Alignment AlgorithmsDPODPO on pairszps.train(dataset, method="dpo")DPO
09 Rejection Samplingrejection samplingbest completion per prompt by reward, with chance controlszps.optimize(rows, mode="sft", select="top_per_prompt", min_reward=0.7)Prune, then publish
10 The Nature of Preferenceshuman preferences, annotatorsjudge checked against hand labelszps.judge_agreement(rows, gold="gold_reward")Trust the numbers
11 Preference Datapreference data, margin, length biaspairs with margin and length gapzps.build_preference_pairs(rows); zps.export_preference(pairs)Training rows
12 Synthetic Data & Distillationsynthetic data, on-policy distillationsimulations; distillation not herezps.simulate(agent, spec=...)Agent simulations
13 Tool Use and Function Callingtool use, function callingtool-call efficiency markerzps.conduct_grade(trajectory)Tool-call efficiency
14 Over-Optimizationover-optimization, reward hackingreward-hack scan plus style drift markerszps.style_report(rows); zps.refusal_report(benign_rows)Trust the numbers
15 RegularizationKL penalty, reference modelKL measured, not appliedzps.mean_kl(rows); zps.logprob_report(rows)RL runs
16 Evaluationevals, contaminationholdout by task, decontamination, before and afterzps.decontaminate(rows, against); zps.delta_report(before, after)Trust the numbers
17 Model Character & Productscharacter training, model speccharacter trainingexamples/characterCharacter training

Calls are import zeroproof.simulations as zps, SDK 0.30. examples/ paths are in the SDK repo. Definitions on the glossary.

Not here

02 A Tiny History of RLHFhistory of RLHF

not here

06 Reinforcement Learningpolicy gradient: PPO, GRPO

GRPO only, no PPO

12 Synthetic Data & Distillationsynthetic data, on-policy distillation

simulations; distillation not here

15 RegularizationKL penalty, reference model

KL measured, not applied

The LLM judge is the default reward; a reward model is one Train press away. No PPO: GRPO and DPO. No on-policy distillation loop. KL is a number on the run page, not a term in the loss; the trainer applies it. Over-optimization is a scan for reward-length and reward-tool-count correlation, not a monitor during training.

Questions, or a higher limit: jacob@zeroproofai.com