# Mission: Understanding LLM Post-Training (Fine-tuning & RL)

## Why
You build software with LLM APIs and agent frameworks, and you want accurate mental
models of how these models are actually made — so the products you ship are better.
Understanding *why* a model behaves the way it does lets you make sharper decisions:
when to just prompt, when fine-tuning would help, which model to reach for, and what
"RLHF" / "alignment" actually buy you. You are **not** training your own models — this
is conceptual fluency in service of building, not a path to becoming an ML engineer.

## Success looks like
- You can explain, unprompted, the difference between **pretraining**, **supervised fine-tuning**, and **preference/RL tuning** — and what each stage gives a model.
- When facing a product problem, you can reason about whether prompting, fine-tuning, or a different model is the right lever (and why).
- You can read a model's release notes / system card and understand the post-training claims (RLHF, DPO, instruction tuning, constitutional AI) without them being jargon.
- You can hold your own in a technical conversation about why models hallucinate, refuse, or "feel aligned."

## Constraints
- **No model training.** Lessons stay conceptual — no GPU, no datasets, no training runs required.
- **Depth: concepts + analogies.** Intuition first, minimal math. Show a number or formula only when it genuinely clarifies.
- **Starting point:** strong software developer, heavy LLM-API user, but fuzzy on internals (training, gradients, loss).
- Lessons should be short and tied to building real AI products.

## Out of scope (for now)
- Hands-on training, fine-tuning code, or framework tutorials (TRL, Axolotl, etc.).
- Deep RL theory for its own sake (Markov decision processes, policy-gradient derivations) beyond what's needed for intuition.
- Pretraining mechanics in depth (tokenization internals, transformer architecture math) — touched only as needed to frame post-training.