Instruction tuning
Fine-tuning a model on many tasks phrased as instructions, so it learns to do what it is asked.
Also: instruction fine-tuning
Instruction tuning trains a model on a wide mix of tasks written as plain-language requests, such as "summarise this article" or "translate into Tamil". Google's FLAN paper in 2021 showed that this made models much better at new instructions they had never seen. OpenAI's InstructGPT, described in 2022, combined it with human feedback. Almost every chat model is now instruction tuned, and open versions often carry "Instruct" or "it" in their names to distinguish them from base models.