Training language models to follow instructions with human feedback

Abstract: Making LLMs bigger does not inherently make them better at following a user's intent. For example, LLMs 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 LLMs 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 LLMs with human intent.

• The paper demonstrates that RLHF fine-tunes LLMs for user alignment, achieving higher human preference even with fewer parameters.
• The methodology uses a three-step process: collecting demonstrations, ranking outputs, and applying PPO-based reinforcement learning.
• Results show marked improvements in truthfulness, reduced toxicity, and maintained NLP performance without sacrificing efficiency.

Training LLMs to Follow Instructions with Human Feedback

Introduction

This paper addresses the alignment issues faced by LLMs like GPT-3 with their users' intentions. Despite being trained on vast web data, these models often exhibit problematic behaviors such as generating untruthful, toxic, or tangential outputs. The discrepancy arises because the pretext task—predicting the next token on internet-derived texts—differs greatly from producing user-aligned responses focused on helpfulness, honesty, and harmlessness. This study introduces InstructGPT, a model fine-tuned using human feedback to better align LLM outputs with human intent across a variety of tasks.

Methodology

The fine-tuning process is a three-step approach leveraging reinforcement learning from human feedback (RLHF). Initially, a set of demonstrations is compiled whereby labelers showcase desired model behaviors. Following this, human rankings of various model outputs reinforce the supervised learning model. Finally, reinforcement learning fine-tunes the model against these rankings using the Proximal Policy Optimization (PPO) algorithm and a reward model predictive of human preferences.

Figure 1: A diagram illustrating the three steps of our method, indicating the use of human feedback in training.

Results and Evaluations

Human evaluations demonstrate significant preference for outputs from InstructGPT over baseline GPT-3 models, despite InstructGPT using fewer parameters. Notably, outputs from a 1.3B parameter InstructGPT model are preferred to those from 175B GPT-3 models. Additionally, the InstructGPT models improve truthfulness and reduce toxic output generation substantially while maintaining comparable performance on standard NLP benchmarks. However, InstructGPT models still produce simple errors occasionally, such as failing to recognize false premises in questions.

Figure 2: Human evaluations of InstructGPT models against GPT-3 baselines, showing preference for InstructGPT outputs on various prompts.

Implications and Open Questions

The findings suggest that RLHF is a promising direction for realigning LLMs with human intentions. By efficiently leveraging human feedback, models can achieve significant alignment without an "alignment tax"—a decrease in performance on established NLP tasks. Future research could explore refining human data collection methodologies, enhancing model responsiveness to adversarial prompts, and conceptualizing models that can adapt to diverse human values and preferences.

Additionally, determining how universally applicable the alignment techniques are, and how they can be adapted to further reduce biased and harmful outputs, remains an open avenue. Work is needed to develop methods ensuring models understand when not to comply with harmful or unethical user requests, which poses a serious challenge in current implementations.

In conclusion, aligning LLMs more closely with human values and preferences is crucial in mitigating risks associated with misuse. The exploration of RLHF offers a viable path towards more reliable and safe AI deployments across various applications, although continued research and ethical considerations remain vital.