Instruction tuning and chat templates
**Instruction tuning** trains models to follow prompts/instructions. **Chat templates** format roles (system/user/assistant) the way a model expects.
What it is
Instruction tuning trains models to follow prompts/instructions. Chat templates format roles (system/user/assistant) the way a model expects.
Why it matters
Wrong templates silently degrade quality. Instruction data choices shape behavior and safety.
How it works (plain)
Base LM predicts web text. Instruction/chat stages teach “answer helpful requests” with special formatting. Always use the model’s official template.
Everyday example
A form that only works if fields are in the right order—chat templates are that form.
Try it
Compare a raw completion prompt vs a proper chat-template prompt on the same model card’s example.
Myths
- ⚠️ Myth: Any “User:” text wrapping is fine.
- ✓ Reality: Special tokens and role layouts are model-specific.
Sources
- Course 07 pretraining-and-finetuning; Course 08 prompting
- Model card chat template sections (cite specifically)