[Paper Review] Pretraining Language Models with Human Preferences

Tomasz Korbak, Kejian Shi+6 more|arXiv (Cornell University)|Feb 16, 2023

Hate Speech and Cyberbullying Detection26 citations

TL;DR

The paper shows that pretraining language models with human-preference objectives (especially conditional training) yields models that significantly reduce undesirable content while preserving downstream performance, outperforming standard MLE pretraining followed by feedback 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.

Motivation & Objective

• Motivate the need to align language models with human preferences during pretraining, not just finetuning.
• Investigate five human-feedback pretraining objectives and compare them to standard MLE pretraining.
• Evaluate alignment and capabilities across toxicity, PII leakage, and PEP8-compliant code tasks.
• Identify Pareto-optimal objectives and provide practical guidance for PHF methods.

Proposed method

• Formulate pretraining with a segment-level reward function R for each training segment.
• Compare five PHF objectives (conditional training, dataset filtering, unlikelihood loss, reward-weighted regression, advantage-weighted regression) against standard MLE.
• Use a 124M parameter GPT-2 small architecture across 3.32B token datasets for toxicity, PII, and PEP8 tasks.
• Evaluate alignment via misalignment scores (negative rewards) and model capabilities via KL divergence from GPT-3 and downstream benchmarks.
• Assess robustness to red-teaming and adversarial prompts across tasks.

Experimental results

Research questions

• RQ1 Can pretraining with human feedback (PHF) outperform standard MLE pretraining followed by feedback finetuning across alignment and capability metrics?
• RQ2 Which PHF objective offers the best alignment-capability trade-off across toxicity, PII, and PEP8 tasks?
• RQ3 Does incorporating human preferences during pretraining create Pareto-optimal solutions for LM safety and usefulness?
• RQ4 How robust are PHF-trained LMs to red-teaming compared to MLE-trained baselines?
• RQ5 Do PHF-trained models retain performance on zero-shot and GLUE-like downstream tasks?

Key findings

• Conditional training is Pareto-optimal across all three tasks, often strictly so for toxicity and PEP8.
• PHF methods reduce undesirable content by up to an order of magnitude compared to standard MLE.
• PHF leads to equal or better alignment than the standard MLE plus finetuning with feedback, and can match zero-shot and GLUE benchmarks.
• Finetuning with feedback from MLE-pretrained models generally underperforms PHF trained from scratch.
• PHF objectives improve adversarial robustness relative to MLE but do not eliminate all vulnerabilities.