Post-training

This optional module sequence follows C1. It uses supervised demonstrations, pairwise preferences, verifiable rewards, and teacher distributions to adapt a small causal language model.

#	Module	Chapter	Kind	Pass
1	L12.1	Supervised fine-tuning: chat rendering, assistant-only loss, packing	build	10
2	L12.2	Direct preference optimization and IPO	build	10
3	L12.3	Group relative policy optimization with verifiable rewards	build	10
4	L12.4	Distillation with forward and reverse KL	build	10