Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

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

TL;DR

The paper tackles the barriers to accessible large language models by addressing high pretraining costs, multilingual performance gaps, and safety/compliance requirements. It introduces Aurora-M, a 15B multilingual continuation of StarCoderPlus trained via a two-stage CAP+CAT curriculum on about 2T tokens and fine-tuned with a Biden-Harris safety instruction corpus. Empirically, Aurora-M demonstrates robust retention of English and code knowledge while achieving strong performance across non-English languages and showing measurable safety improvements through red-teaming. By releasing Aurora-M as an open-source resource, the work advocates responsible, community-driven development of multilingual, safety-aligned LLMs for broad practical impact.

Abstract

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.

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

TL;DR

The paper tackles the barriers to accessible large language models by addressing high pretraining costs, multilingual performance gaps, and safety/compliance requirements. It introduces Aurora-M, a 15B multilingual continuation of StarCoderPlus trained via a two-stage CAP+CAT curriculum on about 2T tokens and fine-tuned with a Biden-Harris safety instruction corpus. Empirically, Aurora-M demonstrates robust retention of English and code knowledge while achieving strong performance across non-English languages and showing measurable safety improvements through red-teaming. By releasing Aurora-M as an open-source resource, the work advocates responsible, community-driven development of multilingual, safety-aligned LLMs for broad practical impact.

Abstract

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.
Paper Structure (37 sections, 6 figures, 7 tables)

This paper contains 37 sections, 6 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Datasets
  3. Data Curation.
  4. Data Filtering.
  5. Data Processing.
  6. Model Training
  7. Safety
  8. Evaluation
  9. Evaluation Setup
  10. Evaluation Results
  11. Evaluation on Natural Languages.
  12. Code Evaluation.
  13. Safety Evaluation
  14. Training Analysis
  15. Related Work
  16. ...and 22 more sections

Figures (6)

  • Figure 1: Comparison of overall performance between Aurora-M-redteamed and its predecessors, StarCoderBase and StarCoderPlus, across diverse code and multilingual language evaluation benchmarks. Pass@1 performance averages for code benchmarks are reported. For natural language evaluations, 0-shot accuracy averages are reported for languages other than English and Japanese. English evaluation is 8-shot, while Japanese evaluation uses a combination of 4-shot and 1-shot.
  • Figure 2: Training data distribution of languages, code, and instructions used for the two-stage continual pretraining of the Aurora-M model. There are a total of 377B and 58B tokens in the Continual Auxiliary Pretraining (CAP) and Continual Alignment Tuning (CAT) stages respectively.
  • Figure 3: Overall safety results.
  • Figure 4: Safety evaluation results comparing our base model and instruction-tuned version.
  • Figure 5: Performance trends of models on HumanEval, MBPP, and English language tasks.
  • ...and 1 more figures