Neura saysInstruction tuning teaches a base model to follow instructions, not just continue text.
A base model completes text; an instruction-tuned model does what you ask. You fine-tune on many (instruction, good-response) examples spanning tasks — summarize, translate, answer, format — so the model generalizes to following new instructions. It's usually the step before preference alignment (RLHF/DPO). Instruction tuning is what makes "just tell the model what to do" work.
Power-ups you unlock
Train on (instruction, response) examples
Generalizes to following new instructions
Bridges base model → usable assistant
Precedes preference alignment
The Overfit Ogre attacks — common mistakes
Expecting a base model to follow instructions
Confusing instruction tuning with RLHF
Narrow instruction data → poor generalization
Boss battleExplain the difference between a base model and an instruction-tuned one.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>base: "The capital of France" → "is a common quiz question..."
instruct-tuned: "What is the capital of France?" → "Paris."</pre></body></html>