Training language models to follow instructions with human feedback
By Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, S. Agarwal, Katarina Slama, Alex Ray, John Schulman and 9 more. Neural Information Processing Systems, 2022
22,660 citations recorded by the source corpus.
Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.
Subfields: Artificial Intelligence, Experimental and Cognitive Psychology, Large Language Models, Alignment & RLHF
View Paper: https://arxiv.org/abs/2203.02155
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Mohit Bansal
- Jianfeng Gao
- Luke S. Zettlemoyer
- Percy Liang
- Quoc V. Le
- Yoshua Bengio
- Yejin Choi
- Graham Neubig
In Experimental and Cognitive Psychology
- Sergey Levine
- Deepak Pathak
- Chelsea Finn
- Percy Liang
- Richard Socher
- Devamanyu Hazarika
- Daniel Klein
- Caiming Xiong
In Large Language Models
- Luke S. Zettlemoyer
- Hinrich Schütze
- Yejin Choi
- Timo Schick
- Noah A. Smith
- Mohit Bansal
- Hannaneh Hajishirzi
- Jianfeng Gao
In Alignment & RLHF
- Angelica Chen
- Geoffrey Irving
- Tomasz Korbak
- Jan Leike
- Ethan Perez
- Richard Ngo
- Danqi Chen
- Sergey Levine
Citations in the corpus
- PaLM: Scaling Language Modeling with Pathways (2022, background): “Finally, additional work has enabled language models to follow instructions (Ouyang et al., 2022; Wei et al., 2022a) – improving the usefulness and reliability of these models.”
- Llama 2: Open Foundation and Fine-Tuned Chat Models (2023, methodology): “As was observed in other works (Stiennon et al., 2020; Ouyang et al., 2022), we find this constraint is useful for training stability, and to reduce reward hackingwherebywewould achieve high scores from the rewardmodel but low scores from human evaluation.”
- Constitutional AI: Harmlessness from AI Feedback (2022, methodology): “D.3 InstructGPT Prompts We warn readers that some of the samples from InstructGPT [Ouyang et al., 2022] were likely cherry picked to display poor model behavior, on purpose, and so may be unrepresentative.”
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model (2023, methodology): “In other words, selecting the model’s desired responses and behavior from its very wide knowledge and abilities is crucial to building AI systems that are safe, performant, and controllable [26].”
- Chain of Thought Prompting Elicits Reasoning in Large Language Models (2022, background): “One recent line of work aims to improve the ability of language models to perform a task by providing instructions that describe the task (Raffel et al., 2020; Wei et al., 2022; Ouyang et al., 2022; Sanh et al., 2022; Wang et al., 2022b).”
- OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalization (2022, background): “These works can be divided into two types, one focusing on fine-tuning the model on a wide range of tasks using human annotated instructions and feedback (Ouyang et al., 2022), whereas the other focusing on adding instructions via annotations or automatically on publicly available benchmarks and dat”
- Emergent Abilities of Large Language Models (2022, methodology): “As another example, Ouyang et al. (2022) proposed a finetuning and reinforcement learning from human feedback approach for the InstructGPT models, which enabled a 1.3B model to outperform much larger models in human-rater evaluations on a broad set of use cases.”
- OPT: Open Pre-trained Transformer Language Models (2022, background): “Future work into instruction learning, in the vein of InstructGPT (Ouyang et al., 2022), may alleviate these limitations.”
- InCoder: A Generative Model for Code Infilling and Synthesis (2022, background): “Moreover, fine-tuning would allow our models to be better able to condition on natural language instructions and other indications of human intent [76, 65, 48].”
- Evaluating the Social Impact of Generative AI Systems in Systems and Society (2023, background): “Addressing this issue may yield improvements in performance for various downstream applications [172].”
- OpenAssistant Conversations - Democratizing Large Language Model Alignment (2023, methodology): “In RLHF , the human acts as a teacher and provides feedback in the form of rewards or penalties.”
- AGIEval: A Human-Centric Benchmark for Evaluating Foundation Models (2023, neutral): “(2022) conducted a holistic evaluation of dozens of language models (up to InstructGPT (Ouyang et al., 2022)) on several previous NLP tasks (e.”
- Self-Refine: Iterative Refinement with Self-Feedback (2023, background): “5 (text-davinci-003 and gpt-3.5-turbo; OpenAI; Ouyang et al., 2022) and GPT-4 (OpenAI, 2023) by 5-40% absolute improvement.”
- Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements (2023, methodology): “( Zhang et al., 2022) proposed the reverse generation method for constructing adversarial data, demonstrating that their data is highly inductive and can more effectively expose safety issues inherent in the models.”
- Augmented Language Models: a Survey (2023, methodology): “Typically, supervised learning is done after an initial pre-training with a language modeling objective (Ouyang et al., 2022; Chung et al., 2022); an exception to this is recent work by Taylor et al.”