Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
Auteurs : Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, Jared Kaplan
Résumé : We apply preference modeling and reinforcement learning from human feedback (RLHF) to finetune language models to act as helpful and harmless assistants. We find this alignment training improves performance on almost all NLP evaluations, and is fully compatible with training for specialized skills such as python coding and summarization. We explore an iterated online mode of training, where preference models and RL policies are updated on a weekly cadence with fresh human feedback data, efficiently improving our datasets and models. Finally, we investigate the robustness of RLHF training, and identify a roughly linear relation between the RL reward and the square root of the KL divergence between the policy and its initialization. Alongside our main results, we perform peripheral analyses on calibration, competing objectives, and the use of OOD detection, compare our models with human writers, and provide samples from our models using prompts appearing in recent related work.
Explorez l'arbre d'article
Cliquez sur les nœuds de l'arborescence pour être redirigé vers un article donné et accéder à leurs résumés et assistant virtuel
Recherchez des articles similaires (en version bêta)
En cliquant sur le bouton ci-dessus, notre algorithme compare le titre et le résumé de cet article à tous les articles de la même catégorie arXiv que nous avons résumés, et trouve les plus proches. Vous pouvez le relancer de temps en temps : les résultats s'améliorent à mesure que notre base grandit.