Base vs instruct vs chat models
The model you actually use is not the raw pretrained one. A second stage — post-training — reshapes a text-completer into an assistant that follows instructions and behaves. Knowing the difference clears up a lot.
Three flavors
| Type | Behaves like | Good for |
|---|---|---|
| Base | autocomplete of the internet | raw generation, research, your own fine-tuning base |
| Instruct | follows a single instruction | one-shot tasks (summarize, classify, extract) |
| Chat | multi-turn conversation with roles | assistants, agents, anything stateful |
Ask a base model "What is the capital of France?" and it might reply with a list of similar quiz questions — it's completing a document, not answering you.
The post-training stack
- SFT (supervised fine-tuning / instruction tuning) — train on curated
(prompt → ideal response)pairs so the model learns the format of being helpful and following instructions. - Preference alignment — RLHF or DPO tunes which of several valid answers the model prefers, using human preference data, to make it more helpful, honest, and harmless.
This is where chat templates, system prompts, and refusal behavior come from. The base model supplies the knowledge; post-training supplies the manners.
Why it matters in practice
- Prompting differs — base models want few-shot examples/continuation; chat models want instructions and a system prompt (prompting).
- Chat templates are real — the special tokens that delimit roles must match the model, or quality drops.
- Alignment is a tax and a gift — it adds safety and usability but can cause over-refusal and a slight capability "alignment tax."
Connects to: pretraining · SFT, RLHF & DPO · prompting base vs chat