Jais 2 is a family of open-weight, Arabic-centric large language models designed for strong Arabic performance and broad cultural coverage. The family includes, to our knowledge, the largest open Arabic-centric LLM trained from scratch, with 70B parameters, alongside an efficient 8B model. A custom Arabic-centric vocabulary supports efficient training and inference, while an optimized architecture and training recipe improve compute efficiency. Jais 2 achieves the highest scores on the Open Arabic LLM Leaderboard v2 (OALL2) and AraGen v1 among the evaluated open models, while remaining competitive on English benchmarks. It also obtains strong results in Arabic poetry, dream interpretation, translation, and summarization. On the IslamicQA domain, Jais 2-70B achieves the second-highest overall average among the evaluated large open models, while attaining the highest scores on PalmX and the held-out in-house IslamicQA benchmark. We release both model sizes as open-weights on Hugging Face under a commercially permissive license. Jais 2-70B is also available as a chat application on the Web, iOS, and Android, reaching inference speeds of up to 2,000 tokens per second on Cerebras hardware.

Large language models remain predominantly English-centric, which limits their utility for underrepresented languages. We help bridge this gap for Hindi with Llama-3-Nanda-10B-Chat (aka Nanda-10B) and Llama-3.1-Nanda-87B-Chat (aka Nanda-87B), forming the Nanda family of open-weight bilingual models (https://github.com/MBZUAI-IFM/Nanda-Family). Our approach integrates: (i) a tokenizer extending Llama’s vocabulary with 20% Hindi-specific tokens, thus halving Hindi tokenization fertility while preserving English efficiency, (ii) Hindi-first parameter-efficient continual pretraining using Llama Pro on a 65B-token corpus spanning Devanagari script, code-mixed, and Romanized Hindi, and (iii) bilingual instruction and safety alignment on a large culturally grounded dataset. The resulting Nanda models outperform open-weight LLMs of comparable size: Nanda-87B yields high generative quality, and Nanda-10B shows competitive general-purpose performance. Nanda-87B demonstrates state-of-the-art performance on summarization, translation, transliteration, and instruction following. Moreover, both models achieve state-of-the-art performance in safety and in cultural knowledge. Our results demonstrate that careful tokenizer design, data curation, and continual pretraining can yield capable and safe LLMs for resource-poor languages without compromising English performance.

We probe pre-trained transformer language models for bridging inference. We first investigate individual attention heads in BERT and observe that attention heads at higher layers prominently focus on bridging relations in-comparison with the lower and middle layers, also, few specific attention heads concentrate consistently on bridging. More importantly, we consider language models as a whole in our second approach where bridging anaphora resolution is formulated as a masked token prediction task (Of-Cloze test). Our formulation produces optimistic results without any fine-tuning, which indicates that pre-trained language models substantially capture bridging inference. Our further investigation shows that the distance between anaphor-antecedent and the context provided to language models play an important role in the inference.

Lexical semantics and world knowledge are crucial for interpreting bridging anaphora. Yet, existing computational methods for acquiring and injecting this type of information into bridging resolution systems suffer important limitations. Based on explicit querying of external knowledge bases, earlier approaches are computationally expensive (hence, hardly scalable) and they map the data to be processed into high-dimensional spaces (careful handling of the curse of dimensionality and overfitting has to be in order). In this work, we take a different and principled approach which naturally addresses these issues. Specifically, we convert the external knowledge source (in this case, WordNet) into a graph, and learn embeddings of the graph nodes of low dimension to capture the crucial features of the graph topology and, at the same time, rich semantic information. Once properly identified from the mention text spans, these low dimensional graph node embeddings are combined with distributional text-based embeddings to provide enhanced mention representations. We illustrate the effectiveness of our approach by evaluating it on commonly used datasets, namely ISNotes and BASHI. Our enhanced mention representations yield significant accuracy improvements on both datasets when compared to different standalone text-based mention representations.

The task of Question Answering is at the very core of machine comprehension. In this paper, we propose a Convolutional Neural Network (CNN) model for text-based multiple choice question answering where questions are based on a particular article. Given an article and a multiple choice question, our model assigns a score to each question-option tuple and chooses the final option accordingly. We test our model on Textbook Question Answering (TQA) and SciQ dataset. Our model outperforms several LSTM-based baseline models on the two datasets.

We introduce a composite deep neural network architecture for supervised and language independent context sensitive lemmatization. The proposed method considers the task as to identify the correct edit tree representing the transformation between a word-lemma pair. To find the lemma of a surface word, we exploit two successive bidirectional gated recurrent structures - the first one is used to extract the character level dependencies and the next one captures the contextual information of the given word. The key advantages of our model compared to the state-of-the-art lemmatizers such as Lemming and Morfette are - (i) it is independent of human decided features (ii) except the gold lemma, no other expensive morphological attribute is required for joint learning. We evaluate the lemmatizer on nine languages - Bengali, Catalan, Dutch, Hindi, Hungarian, Italian, Latin, Romanian and Spanish. It is found that except Bengali, the proposed method outperforms Lemming and Morfette on the other languages. To train the model on Bengali, we develop a gold lemma annotated dataset (having 1,702 sentences with a total of 20,257 word tokens), which is an additional contribution of this work.