Glossary
Instruction data
Datasets pairing instructions or prompts with high-quality responses, used in supervised fine-tuning to teach a model to follow directions.
Each example pairs a prompt with a demonstration of the desired response, often with system context or multi-turn structure. Quality is judged on correctness, helpfulness, and diversity of tasks; research has shown that a small set of excellent examples can rival much larger mediocre ones. Expert-written examples in specialist domains — medicine, law, code — are the scarce end.
Why it matters
Instruction data is valued by the example, and the value tracks the expertise behind each one rather than raw volume.