Dongxi NLP explains KL divergence for post-training models

Dongxi NLP publishes an explanation of Kullback–Leibler (KL) divergence as the sixth part of its post-training series.

The article describes KL as a measure of how far a fine-tuned language model's output distribution has moved from its reference model, and it uses a boiled-egg example in which probabilities shift from 0.40 to 0.30 and from 0.12 to 0.22, giving a KL of 0.047 nats. The author says KL is never negative and that it serves as the penalty that keeps reinforcement-learning training close to the reference model. The post says only multiplication skills are needed to follow it.