reinforcement learning from human feedback
ree-IN-fors-ment LUR-ning from HYOO-mun FEED-bak
ˌriːɪnˈfɔːrsmənt ˈlɜːrnɪŋ frɒm ˈhjuːmən ˈfiːdbækListen
Hear it in a clear American voice. Slow it down to catch each syllable.
Syllable breakdown
Dictionary meaning
a fine-tuning technique in which a language model's outputs are rated by human evaluators for qualities such as helpfulness, harmlessness, and honesty, and these ratings are used to train a reward model whose signal guides further training of the language model via reinforcement learning; RLHF was central to the development of InstructGPT and subsequently ChatGPT and has been widely adopted as the primary method for aligning large language models with human preferences; critics note that RLHF's dependence on human rater judgements means that models trained this way can reflect the biases and blind spots of their rater pools
Learn one new pronunciation a day
Get a word, name or idiom like “reinforcement learning from human feedback” in your inbox — daily or weekly, free. Unsubscribe any time.