arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2406.02886cs.CLcs.AI
收藏

PLaD:基于伪偏好对的偏好驱动大语言模型蒸馏

PLaD: Preference-based Large Language Model Distillation with Pseudo-Preference Pairs

Georgia Institute of Technology(佐治亚理工学院);Google(谷歌公司);Google DeepMind(谷歌DeepMind)

Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Haorui Wang, Zhen Qin, Feng Han, Jialu Liu, Simon Baumgartner, Michael Bendersky, Chao Zhang

AI总结:

PLaD框架利用师生容量差异生成伪偏好对,通过排序损失重新校准学生模型对输出相对质量的估计,从而在无需访问教师内部状态的情况下实现高效的大语言模型蒸馏。

摘要:

大语言模型(LLMs)在各类任务中展现出卓越能力,但其庞大的参数规模限制了它们在资源受限环境中的适用性。知识蒸馏(KD)通过将大型教师模型的专业知识迁移到紧凑的学生模型,提供了一种可行的解决方案。然而,传统知识蒸馏技术在应用于大语言模型时面临特定挑战,包括对LLM输出的访问受限、教师与学生之间显著的容量差距,以及继承的误校准问题。在本工作中,我们提出了PLaD,一种新颖的基于偏好的大语言模型蒸馏框架。PLaD利用教师与学生之间的容量差异生成伪偏好对,其中教师输出被偏好于学生输出。然后,PLaD利用排序损失重新校准学生对序列似然的估计,引导学生关注理解输出的相对质量,而不是简单地模仿教师。PLaD绕过了访问教师LLM内部状态的需求,解决了学生的表达能力限制,并缓解了学生的误校准问题。通过在两个序列生成任务上使用多种大语言模型进行的大量实验,我们证明了所提出的PLaD框架的有效性。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑