XNews

First Systematic Book on RLHF: Covers DPO, RLVR, and Data-Centric Insights

The first systematic book on RLHF, authored by Nathan Lambert, covers the main pipeline from instruction tuning to reward modeling and reinforcement learning, including DPO, RLVR, synthetic data, over-optimization, and model personality shaping. Lambert's core argument is that data is far more important than algorithms. The book is available online with open-source code.

Source posts