Instruction tuning

Fine-tuning a language model on many different tasks, each written as a plain-language instruction with its answer, so that the model learns to follow instructions for tasks it has not seen.