Open textbook · English / 简体中文
Post-Training of
Large Language Models
Alignment, Reasoning, and Agency
A graduate textbook covering the foundations and modern practice of large-language-model post-training, from supervised fine-tuning and preference alignment to reasoning, tool use, agentic reinforcement learning, systems, evaluation, and safety.
Star this textbook on GitHubSix parts · Twenty-six chapters
From alignment foundations to agentic systems
-
I
Foundations
Post-training paradigms, mathematical preliminaries, SFT, and data.
-
II
Alignment from feedback
Reward models, policy gradients, RLHF, DPO, and online alignment.
-
III
Reasoning
Test-time compute, RLVR, group-relative methods, and credit assignment.
-
IV
Agency
Tool-integrated RL, long-horizon training, environments, and self-play.
-
V
Systems
Training at scale, rollout mismatch, efficiency, and cost control.
-
VI
Measurement and frontiers
Evaluation, safety alignment, robustness, and open problems.