Under-represented languages suffer from a lack of data, and as a result, there are few LLMs that support them. Extending an existing LLM to a new language is a practical option for startups, university labs, and organizations with limited budgets. This process involves several steps. In this paper, we describe how we adapted the Falcon3-7B model to Arabic, covering everything from data collection and training to evaluation. Falcon-Arabic was trained exclusively on native data to better capture the cultural and linguistic aspects of the language. Our evaluations show that Falcon-Arabic achieves state-of-the-art results on a range of Arabic benchmarks.

Arabic is one of the most widely spoken languages in the world, yet efforts to develop and evaluate Large Language Models (LLMs) for Arabic remain relatively limited. Most existing Arabic benchmarks focus on linguistic, cultural, or religious content, leaving a significant gap in areas like STEM and coding domains that are increasingly relevant for real-world LLM applications. To help bridge this gap, we present 3LM, a suite of three benchmarks designed specifically for Arabic. The first is a set of STEM-related question-answer pairs, naturally sourced from Arabic textbooks and educational worksheets. The second consists of synthetically generated STEM questions, created using the same sources. The third benchmark focuses on code generation, built through a careful translation of two widely used code benchmarks, incorporating a human-in-the-loop process with several rounds of review to ensure high-quality and faithful translations. We release all three benchmarks publicly to support the growth of Arabic LLM research in these essential but underrepresented areas.

Construire des systèmes de dialogue qui conversent avec les humains afin de les aider dans leurs tâches quotidiennes est devenu une priorité. Certains de ces systèmes produisent des dialogues en cherchant le meilleur énoncé (réponse) parmi un ensemble d’énoncés candidats. Le choix de la réponse est conditionné par l’historique de la conversation appelé contexte. Ces systèmes ordonnent les énoncés candidats par leur adéquation au contexte, le meilleur est ensuite choisi. Les approches existantes à base de réseaux de neurones profonds sont performantes pour cette tâche. Dans cet article, nous améliorons une approche état de l’art à base d’un dual encodeur LSTM. En se basant sur la similarité sémantique entre le contexte et la réponse, notre approche apprend à mieux distinguer les bonnes réponses des mauvaises. Les résultats expérimentaux sur un large corpus de chats d’Ubuntu montrent une amélioration significative de 7, 6 et 2 points sur le Rappel@(1, 2 et 5) respectivement par rapport au meilleur système état de l’art.

Since the advent of word embedding methods, the representation of longer pieces of texts such as sentences and paragraphs is gaining more and more interest, especially for textual similarity tasks. Mikolov et al. (2013) have demonstrated that words and phrases exhibit linear structures that allow to meaningfully combine words by an element-wise addition of their vector representations. Recently, Arora et al. (2017) have shown that removing the projections of the weighted average sum of word embedding vectors on their first principal components, outperforms sophisticated supervised methods including RNN’s and LSTM’s. Inspired by Mikolov et al. (2013) and Arora et al. (2017) findings and by a bilingual word mapping technique presented in Artetxe et al. (2016), we introduce MappSent, a novel approach for textual similarity. Based on a linear sentence embedding representation, its principle is to build a matrix that maps sentences in a joint-subspace where similar sets of sentences are pushed closer. We evaluate our approach on the SemEval 2016/2017 question-to-question similarity task and show that overall MappSent achieves competitive results and outperforms in most cases state-of-art methods.