arXiv 2404.003992024-12-27cs.CLcs.AIcs.LG

Aurora-M：面向多语言与代码的开源持续预训练

Aurora-M: Open Source Continual Pre-training for Multilingual Language and Code

Babelscape（巴别景观公司）；LAION（LAION组织）；TU Darmstadt（达姆施塔特工业大学）；UNC Chapel-Hill（北卡罗来纳大学教堂山分校）；Detomo Inc.（Detomo公司）；CSIRO’s Data61（联邦科学与工业研究组织Data61）；Monash University（莫纳什大学）；Mila - Quebec AI Institute（米拉-魁北克人工智能研究所）；Carnegie Mellon University（卡内基梅隆大学）；Allen Institute for AI（艾伦人工智能研究所）；DeepTensor AB（DeepTensor AB公司）；WASP Media & Language（WASP媒体与语言机构）；Umeå University（于默奥大学）；University of Massachusetts Amherst（马萨诸塞大学阿默斯特分校）；TU Wien（维也纳技术大学）；Institute of Science Tokyo（东京科学大学）；Hudson Labs（哈德逊实验室）；Luleå University of Technology（吕勒奥理工大学）；ETH Zurich（苏黎世联邦理工学院）；RunPod（RunPod公司）；OpenChat（OpenChat机构）；RIKEN CCS（理化学研究所计算科学中心）；Chattermill AI（Chattermill人工智能公司）；ASC27（ASC27机构）；University of Massachuset（马萨诸塞大学）

Taishi Nakamura, Mayank Mishra, Simone Tedeschi, Yekun Chai, Jason T Stillerman, Felix Friedrich, Prateek Yadav, Tanmay Laud, Vu Minh Chien, Terry Yue Zhuo, Dig… 展开作者

Taishi Nakamura, Mayank Mishra, Simone Tedeschi, Yekun Chai, Jason T Stillerman, Felix Friedrich, Prateek Yadav, Tanmay Laud, Vu Minh Chien, Terry Yue Zhuo, Diganta Misra, Ben Bogin, Xuan-Son Vu, Marzena Karpinska, Arnav Varma Dantuluri, Wojciech Kusa, Tommaso Furlanello, Rio Yokota, Niklas Muennighoff, Suhas Pai, Tosin Adewumi, Veronika Laippala, Xiaozhe Yao, Adalberto Junior, Alpay Ariyak, Aleksandr Drozd, Jordan Clive, Kshitij Gupta, Liangyu Chen, Qi Sun, Ken Tsui, Noah Persaud, Nour Fahmy, Tianlong Chen, Mohit Bansal, Nicolo Monti, Tai Dang, Ziyang Luo, Tien-Tung Bui, Roberto Navigli, Virendra Mehta, Matthew Blumberg, Victor May, Huu Nguyen, Sampo Pyysalo

备注：: Preprint

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

Aurora-M 是一个从 StarCoderPlus 持续预训练而来的 150 亿参数多语言开源模型，覆盖六种语言与代码，总训练 token 超 2 万亿，并在人工审核的安全指令上微调，显著提升了多语言性能与安全性。

摘要：

预训练语言模型是人工智能应用的重要组成部分，但其训练所需的高昂计算成本限制了可及性。Bloom 和 StarCoder 等倡议旨在使预训练模型的获取民主化，以促进社区协作开发。尽管做出了这些努力，此类模型仍面临诸多挑战，例如多语言能力有限、持续预训练过程中存在灾难性遗忘风险、从头训练模型成本高昂，以及需要符合人工智能安全标准和监管框架。本文提出了 Aurora-M，这是一个拥有 150 亿参数的多语言开源模型，在英语、芬兰语、印地语、日语、越南语和代码上进行了训练。Aurora-M 从 StarCoderPlus 出发，在额外的 4350 亿个 token 上进行了持续预训练，使其总训练 token 数超过 2 万亿。它是首个在经人工审核的安全指令上进行微调的开源多语言模型，因此其开发不仅符合传统的红队测试考量，还契合拜登-哈里斯《关于安全、可靠和可信地开发和使用人工智能的行政命令》中所阐述的具体关切。我们在广泛的任务和语言上对 Aurora-M 进行了评估，展示了其对灾难性遗忘的鲁棒性以及在多语言环境中的卓越性能，尤其是在安全性评估方面。我们开源了 Aurora-M 及其变体，以鼓励负责任地开源开发大型语言模型，地址为 https://huggingface.co/aurora-m。

英文摘要：

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

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。