arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.02791cs.CLcs.AIcs.LG
收藏

PanGu-$π$ Pro:重新思考微型语言模型的优化与架构

PanGu-$π$ Pro:Rethinking Optimization and Architecture for Tiny Language Models

Huawei Noah’s Ark Lab(华为诺亚方舟实验室);Peking University(北京大学);Consumer Business Group, Huawei(华为消费者业务集团);Huawei Kirin Solution(华为麒麟解决方案部门)

Yehui Tang, Kai Han, Fangcheng Liu, Yunsheng Ni, Yuchuan Tian, Zheyuan Bai, Yi-Qi Hu, Sichao Liu, Shangling Jui, Yunhe Wang

AI总结:

针对移动端微型语言模型的计算与内存挑战,本文基于1B参数模型从架构、初始化和优化策略三方面进行实证研究,提出分词器压缩等有效设计公式,训练的PanGu-$π$ Pro模型在基准上显著提升并超越更大尺寸的SOTA模型。

摘要:

大型语言模型(LLMs)的能力已通过海量数据和计算资源得到证明。然而,在移动设备上应用语言模型面临着计算和内存成本的巨大挑战,即迫切需要高性能的微型语言模型。受限于高度复杂的训练过程,优化语言模型的许多细节很少得到仔细研究。在本研究中,基于具有1B参数的微型语言模型,我们精心设计了一系列实证研究来分析各个组件的效果。主要讨论了三个视角,即神经架构、参数初始化和优化策略。几种设计公式被实证证明对微型语言模型特别有效,包括分词器压缩、架构微调、参数继承和多次训练。然后,我们遵循既定公式,在1.6T多语言语料库上训练了PanGu-$π$-1B Pro和PanGu-$π$-1.5B Pro。实验结果表明,改进的优化和架构使PanGu-$π$-1B Pro在基准评估集上获得了8.87的显著平均提升。此外,PanGu-$π$-1.5B Pro超越了一系列具有更大模型尺寸的SOTA模型,验证了其卓越的性能。代码可在https://github.com/YuchuanTian/RethinkTinyLM获取。

英文摘要:

The power of large language models (LLMs) has been demonstrated through numerous data and computing resources. However, the application of language models on mobile devices is facing huge challenge on the computation and memory costs, that is, tiny language models with high performance are urgently required. Limited by the highly complex training process, there are many details for optimizing language models that are seldom studied carefully. In this study, based on a tiny language model with 1B parameters, we carefully design a series of empirical study to analyze the effect of each component. Three perspectives are mainly discussed, \ie, neural architecture, parameter initialization, and optimization strategy. Several design formulas are empirically proved especially effective for tiny language models, including tokenizer compression, architecture tweaking, parameter inheritance and multiple-round training. Then we train PanGu-$π$-1B Pro and PanGu-$π$-1.5B Pro on 1.6T multilingual corpora, following the established formulas. Experimental results demonstrate the improved optimization and architecture yield a notable average improvement of 8.87 on benchmark evaluation sets for PanGu-$π$-1B Pro. Besides, PanGu-$π$-1.5B Pro surpasses a range of SOTA models with larger model sizes, validating its superior performance. The code is available at https://github.com/YuchuanTian/RethinkTinyLM.

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。

↑