Pretrained language models are 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.

Rhetorical Roles (RR) prediction is to predict the label of each sentence in legal documents, which is regarded as an emergent task for legal document understanding. In this study, we present a novel method for the RR task by exploiting the long context representation. Specifically, legal documents are known as long texts, in which previous works have no ability to consider the inherent dependencies among sentences. In this paper, we propose GNNRR (Graph Neural Network for Rhetorical Roles Prediction), which is able to model the cross-information for long texts. Furthermore, we develop multitask learning by incorporating label shift prediction (LSP) for segmenting a legal document. The proposed model is evaluated on the SemEval 2023 Task 6 - Legal Eval Understanding Legal Texts for RR sub-task. Accordingly, our method achieves the top 4 in the public leaderboard of the sub-task. Our source code is available for further investigation\footnote{https://github.com/hiepnh137/SemEval2023-Task6-Rhetorical-Roles}.