RLTHF: Targeted Human Feedback for LLM Alignment

TL;DR

RLTHF is a hybrid framework that efficiently aligns large language models with human preferences by combining initial LLM-based alignment with targeted human corrections, significantly reducing human effort while maintaining high quality.

Contribution

RLTHF introduces a novel approach that combines LLMs and selective human annotations to achieve full-human alignment with minimal human effort.

Findings

RLTHF achieves full-human annotation-level alignment with only 6-7% of human effort.

Models trained on RLTHF-curated data outperform those trained on fully human-annotated datasets.

RLTHF effectively identifies hard-to-annotate samples for targeted human correction.

Abstract

Fine-tuning large language models (LLMs) to align with user preferences is challenging due to the high cost of quality human annotations in Reinforcement Learning from Human Feedback (RLHF) and the generalizability limitations of AI Feedback. To address these challenges, we propose RLTHF, a human-AI hybrid framework that combines LLM-based initial alignment with selective human annotations to achieve full-human annotation alignment with minimal effort. RLTHF identifies hard-to-annotate samples mislabeled by LLMs using a reward model's reward distribution and iteratively enhances alignment by integrating strategic human corrections while leveraging LLM's correctly labeled samples. Evaluations on HH-RLHF and TL;DR datasets show that RLTHF reaches full-human annotation-level alignment with only 6-7% of the human annotation effort. Furthermore, models trained on RLTHF's curated datasets for downstream tasks outperform those trained on fully human-annotated datasets, underscoring the effectiveness of RLTHF.