Instruction-Following
The ability of a language model to accurately follow user instructions and respond helpfully.
The ability of a language model to accurately follow user instructions and respond helpfully. Emerges from RLHF training on preferences that reward clarity, relevance, and instruction adherence. InstructGPT demonstrates strong instruction-following while GPT-3 often ignores or misinterprets instructions.
How reliably a model does exactly what the prompt asks. Clear, specific, ordered instructions improve it.