LESSON 5.1 · Intuition · 80 min
The gap between a base model and an assistant
The gap between a base model and an assistant. Learn pretraining distributions, chat templates, and behavior shaping and complete: Compare base and instruct output distributions. P
Learning objectives
- Explain base versus instruct outputs through training distributions.
- Show how a chat template changes token sequences and role boundaries.
- Separate SFT behavior shaping from adding factual knowledge.
- Build an auditable paired comparison across six prompt types.
Core concepts
Pretraining distribution
A base model optimizes next-token likelihood over web text. Questions, answers, code, and dialogue may appear without one universal assistant protocol.
Base-model behavior
A question can be answered or continued as a webpage. The form follows the training distribution and prompt prefix; it is not a direct knowledge-capability switch.
Chat template
Templates encode system, user, and assistant roles with special tokens. A tokenizer/template mismatch can cause role leakage, gibberish, or nontermination.
SFT demonstrations
SFT raises the probability of demonstrated assistant responses. It shapes behavior but does not guarantee factuality, calibrated refusal, or every format constraint.
Assistant behavior
Stopping, tone, format, and tool protocols are learned behaviors layered on next-token generation.
Build and verify
Run a six-class base/instruct paired experiment
- Choose factual, format, multi-turn, uncertainty, instruction-following, and webpage-continuation prompts.
- Save raw base tokens and apply_chat_template tokens.
- Fix temperature, top-p, seed, and length; record output, stopping, and roles.
- Explain each row by training distribution; label template failures separately.