Section 2

Training and Alignment

Llama 3.1 8B and Llama 3.1 8B Instruct share an architecture and 8,030,261,248 parameters; post-training alone separates them. This section follows the Llama 3.1 family from pretraining, planned with 6ND and scaling laws, through supervised and low-rank fine-tuning and preference optimization with PPO, DPO and GRPO, to the specifications, refusal measurements and classifiers of a production alignment stack.