daDPO: Distribution-Aware DPO for Distilling Conversational Abilities

TL;DR

This paper introduces daDPO, a novel distribution-aware preference optimization method that improves the conversational abilities of smaller language models through better distillation, outperforming existing approaches in empirical tests.

Contribution

The paper proposes daDPO, a unified framework combining preference optimization and distribution-based distillation, with theoretical analysis and empirical validation showing superior performance.

Findings

20% pruned Vicuna1.5-7B achieves near-teacher performance

Qwen2.5-1.5B outperforms its 7B teacher model in some cases

daDPO outperforms existing methods in restoring and enhancing model performance

Abstract

Large language models (LLMs) have demonstrated exceptional performance across various applications, but their conversational abilities decline sharply as model size decreases, presenting a barrier to their deployment in resource-constrained environments. Knowledge distillation with Direct Preference Optimization (dDPO) has emerged as a promising approach to enhancing the conversational abilities of smaller models using a larger teacher model. However, current methods primarily focus on 'black-box' KD, which only uses the teacher's responses, overlooking the output distribution offered by the teacher. This paper addresses this gap by introducing daDPO (Distribution-Aware DPO), a unified method for preference optimization and distribution-based distillation. We provide rigorous theoretical analysis and empirical validation, showing that daDPO outperforms existing methods in restoring performance for pruned models and enhancing smaller LLM models. Notably, in in-domain evaluation, our method enables a 20% pruned Vicuna1.5-7B to achieve near-teacher performance (-7.3% preference rate compared to that of dDPO's -31%), and allows Qwen2.5-1.5B to occasionally outperform its 7B teacher model (14.0% win rate).