Pretraining Language Models with Human Preferences

TL;DR

This paper investigates pretraining language models with human preferences using various objectives, finding that conditional training effectively reduces undesirable content while maintaining task performance, suggesting a shift beyond imitation learning.

Contribution

It introduces and benchmarks a simple conditional training approach for pretraining LMs with human feedback, improving alignment without sacrificing capabilities.

Findings

Conditional training reduces undesirable content by up to tenfold.

Pretraining with human preferences outperforms standard methods followed by fine-tuning.

The approach maintains downstream task performance after finetuning.

Abstract

Language models (LMs) are pretrained to imitate internet text, including content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, and more. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the trade-off between alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores given by a reward model. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.