Training / Standard term
Instruction tuning
Fine-tuning on instruction-response examples so a model more often maps natural-language tasks to intended outputs.
Examples can cover questions, transformations, tool use, formats, dialogue, and refusals. They may be written by people, generated by models, or filtered from other data. Instruction tuning changes response behavior but does not stop next-token prediction from being the underlying training and generation mechanism.
Builder example
Instruction-tuned variants are often a strong starting point for assistant interfaces, while base models may suit research or specialized training pipelines. Labels such as chat and instruct do not establish quality on a particular schema or workflow, so test the exact variant.
Common confusion: Instruction tuning and preference optimization can occur in several orders or be combined. Some pipelines use reinforcement learning from human feedback (RLHF), while others use different preference methods or no separate preference stage.