PLaD: Preference-based Large Language Model Distillation with Pseudo-Preference Pairs
This work addresses a critical bottleneck in deploying LLMs efficiently, offering a novel distillation method that is incremental but impactful for applications requiring compact models.
The paper tackles the challenge of distilling large language models (LLMs) into smaller ones for resource-constrained settings by introducing PLaD, a preference-based framework that uses pseudo-preference pairs and ranking loss to improve student model calibration and performance, achieving competitive results on sequence generation tasks.
Large Language Models (LLMs) have exhibited impressive capabilities in various tasks, yet their vast parameter sizes restrict their applicability in resource-constrained settings. Knowledge distillation (KD) offers a viable solution by transferring expertise from large teacher models to compact student models. However, traditional KD techniques face specific challenges when applied to LLMs, including restricted access to LLM outputs, significant teacher-student capacity gaps, and the inherited mis-calibration issue. In this work, we present PLaD, a novel preference-based LLM distillation framework. PLaD exploits the teacher-student capacity discrepancy to generate pseudo-preference pairs where teacher outputs are preferred over student outputs. Then, PLaD leverages a ranking loss to re-calibrate student's estimation of sequence likelihood, which steers the student's focus towards understanding the relative quality of outputs instead of simply imitating the teacher. PLaD bypasses the need for access to teacher LLM's internal states, tackles the student's expressivity limitations, and mitigates the student mis-calibration issue. Through extensive experiments on two sequence generation tasks and with various LLMs, we demonstrate the effectiveness of our proposed PLaD framework.