Instruction Following › Instruction Following

Imagine you've just finished pre-training a large language model on trillions of tokens from the internet. The model can complete sentences, generate coherent text, and even demonstrate surprising capabilities like answering questions when prompted correctly. But when you ask it to "Summarize this article in three bullet points," it continues writing more of the article instead. When you request "Translate the following sentence to French," it sometimes translates and sometimes just keeps generating English. The model has learned language, but it hasn't learned to follow instructions.

This gap between language modeling capability and practical utility motivated one of the most important developments in modern LLMs: instruction tuning. As we explored in Part XXVIII with GPT-3's in-context learning, pre-trained models can perform tasks when given the right examples in their context. But this requires careful prompt engineering, wastes context space on demonstrations, and fails unpredictably on novel phrasings. Instruction tuning offers a different approach: teaching models to understand and execute your requests directly, without requiring examples at inference time.

Think of the gap between a base language model and an instruction-following assistant like the difference between a brilliant but socially oblivious expert and a skilled consultant. The expert possesses all the knowledge and reasoning ability you need, but they don't recognize when you're asking them for help versus when you're just making conversation. The consultant has learned a social contract: when someone presents a problem, they engage with it directly, provide a concrete recommendation, and stop when they've addressed what was asked. Instruction tuning teaches language models this same social contract, transforming their raw capability into practical utility.

Understanding why this transformation is necessary, how the training process achieves it, and what factors determine its success requires examining the basic nature of language model pre-training and its relationship to the tasks you want models to perform. The mismatch isn't a bug in the pre-training procedure; it's an inevitable consequence of optimizing for one objective while wanting behavior aligned with a different one. Recognizing this helps you understand what instruction tuning teaches and why certain design choices in instruction data matter so much.

This chapter introduces the motivation behind instruction tuning, explores how to design effective instruction formats, examines why diversity in training data matters for generalization, and discusses what distinguishes high-quality instruction data from noise. Understanding these foundations prepares you for the subsequent chapters on data creation, training procedures, and evaluation methods. You'll also encounter the key concepts that recur throughout the field: the pre-training gap, format consistency, diversity as a proxy for generalization, and the tension between data quality and scale.