Reinforcement Learning from Human Feedback

A training procedure that uses human-rated examples to teach a language model which responses are preferred. The pipeline typically trains a reward model on those ratings, then uses reinforcement learning to update the language model so it earns higher reward. RLHF is the technique that turned base GPT-style models into helpful assistants.

Definición editorial de DecipherU. Para definiciones técnicas autorizadas, consulta NIST AI RMF, ISO/IEC 42001, o las publicaciones oficiales de los laboratorios de IA.