llm-training-pipeline-stages

IN premise — entries/2026/06/21/wiki-LLaMA-chunk-4.md

Created 2026-06-21T09:50:09+00:00

The standard LLM training pipeline proceeds through self-supervised pretraining, then supervised fine-tuning/instruction tuning, then alignment via RLHF or Constitutional AI.

Summary

Most large language models are built in three sequential stages: first learning general language patterns from raw text, then learning to follow instructions, and finally being steered to behave helpfully and safely. This layered structure matters because each stage builds on the one before it, so flaws, biases, or assumptions baked in early carry forward and shape the final system in ways that are difficult to correct afterward.

Dependents

These beliefs depend on this one:

IN training-pipeline-is-mature-engineering-discipline — The LLM training pipeline has matured into a fully specified engineering discipline with standardized stages (pretraining → SFT → alignment), complete mathematical specification of the alignment phase (Bradley-Terry loss, KL penalty, PPO clipping), and calibrated cost models (6 FLOPs/param/token) enabling predictable budgeting from BERT's $500 to PaLM's millions.