Module 5 · Lesson 11

A raw language model and a polished conversational assistant are not the same thing.

Pre-training gives a model broad language capabilities by training it to predict tokens across a large corpus. Post-training then shapes how that capability is expressed when people interact with the model.

What pre-training produces

A base model learns a probability distribution over token sequences. Given text, it predicts plausible continuations. That behavior is probabilistic, but it is not “random completion.” The probabilities reflect patterns learned during training.

A base model may continue a question as if it were part of a document instead of treating it as a request from a user. That is one reason a strong base model can still feel awkward as an assistant.

Supervised fine-tuning teaches interaction patterns

One common post-training step is supervised fine-tuning, often abbreviated SFT. The model trains on carefully constructed examples of prompts and desirable responses.

Base model
   ↓
instruction / supervised fine-tuning
   ↓
preference and safety shaping
   ↓
assistant model

These examples can teach formats such as question answering, summarization, tool-use conventions, structured output, and conversational instruction following.

Post-training is broader than one technique

Modern post-training may combine SFT, preference learning, reinforcement learning, direct preference optimization, AI-generated feedback, safety data, tool-use training, and other methods.

The exact recipe differs across model families. The general purpose is easier to remember: pre-training builds broad capability; post-training shapes behavior for intended use.

Does post-training provide all the model's factual knowledge?

Usually not. Most broad knowledge is learned during pre-training. Post-training can add task-specific information or examples, but it is generally not the most efficient way to keep a model synchronized with changing facts.

For frequently changing or private knowledge, retrieval techniques are often a better fit. We will get to that in the RAG lesson.

Thinking in stages also helps explain why two models with similar base capabilities can behave very differently after post-training. The assistant you experience is the result of both the underlying model and the behavior-shaping process applied afterward.

Demystifying AI series

← Why LLM Pre-Training Uses So Many Tokens   ·   RLHF and RLAIF →