Sunil Kumar Sahu

Neha Sengupta, Sunil Kumar Sahu, Bokang Jia et al. · berkeley

We introduce Jais and Jais-chat, new state-of-the-art Arabic-centric foundation and instruction-tuned open generative large language models (LLMs). The models are based on the GPT-3 decoder-only architecture and are pretrained on a mixture of Arabic and English texts, including source code in various programming languages. With 13 billion parameters, they demonstrate better knowledge and reasoning capabilities in Arabic than any existing open Arabic and multilingual models by a sizable margin, based on extensive evaluation. Moreover, the models are competitive in English compared to English-centric open models of similar size, despite being trained on much less English data. We provide a detailed description of the training, the tuning, the safety alignment, and the evaluation of the models. We release two open versions of the model -- the foundation Jais model, and an instruction-tuned Jais-chat variant -- with the aim of promoting research on Arabic LLMs. Available at https://huggingface.co/inception-mbzuai/jais-13b-chat

Gurpreet Gosal, Yishi Xu, Gokul Ramakrishnan et al.

We present an efficient method for adapting a monolingual Large Language Model (LLM) to another language, addressing challenges of catastrophic forgetting and tokenizer limitations. We focus this study on adapting Llama 2 to Arabic. Our two-stage approach begins with expanding the vocabulary and training only the embeddings matrix, followed by full model continual pre-training on a bilingual corpus. By continually pre-training on a mix of Arabic and English corpora, the model retains its proficiency in English while acquiring capabilities in Arabic. Our approach results in significant improvements in Arabic and slight enhancements in English, demonstrating cost-effective cross-lingual transfer. We perform ablations on embedding initialization techniques, data mix ratios, and learning rates and release a detailed training recipe. To demonstrate generalizability of this approach we also adapted Llama 3 8B to Arabic and Llama 2 13B to Hindi.

Monojit Choudhury, Shivam Chauhan, Rocktim Jyoti Das et al.

Developing high-quality large language models (LLMs) for moderately resourced languages presents unique challenges in data availability, model adaptation, and evaluation. We introduce Llama-3-Nanda-10B-Chat, or Nanda for short, a state-of-the-art Hindi-centric instruction-tuned generative LLM, designed to push the boundaries of open-source Hindi language models. Built upon Llama-3-8B, Nanda incorporates continuous pre-training with expanded transformer blocks, leveraging the Llama Pro methodology. A key challenge was the limited availability of high-quality Hindi text data; we addressed this through rigorous data curation, augmentation, and strategic bilingual training, balancing Hindi and English corpora to optimize cross-linguistic knowledge transfer. With 10 billion parameters, Nanda stands among the top-performing open-source Hindi and multilingual models of similar scale, demonstrating significant advantages over many existing models. We provide an in-depth discussion of training strategies, fine-tuning techniques, safety alignment, and evaluation metrics, demonstrating how these approaches enabled Nanda to achieve state-of-the-art results. By open-sourcing Nanda, we aim to advance research in Hindi LLMs and support a wide range of real-world applications across academia, industry, and public services.

Sunil Kumar Sahu, Derek Thomas, Billy Chiu et al.

Relation Extraction is a way of obtaining the semantic relationship between entities in text. The state-of-the-art methods use linguistic tools to build a graph for the text in which the entities appear and then a Graph Convolutional Network (GCN) is employed to encode the pre-built graphs. Although their performance is promising, the reliance on linguistic tools results in a non end-to-end process. In this work, we propose a novel model, the Self-determined Graph Convolutional Network (SGCN), which determines a weighted graph using a self-attention mechanism, rather using any linguistic tool. Then, the self-determined graph is encoded using a GCN. We test our model on the TACRED dataset and achieve the state-of-the-art result. Our experiments show that SGCN outperforms the traditional GCN, which uses dependency parsing tools to build the graph.

Sunil Kumar Sahu, Fenia Christopoulou, Makoto Miwa et al.

Inter-sentence relation extraction deals with a number of complex semantic relationships in documents, which require local, non-local, syntactic and semantic dependencies. Existing methods do not fully exploit such dependencies. We present a novel inter-sentence relation extraction model that builds a labelled edge graph convolutional neural network model on a document-level graph. The graph is constructed using various inter- and intra-sentence dependencies to capture local and non-local dependency information. In order to predict the relation of an entity pair, we utilise multi-instance learning with bi-affine pairwise scoring. Experimental results show that our model achieves comparable performance to the state-of-the-art neural models on two biochemistry datasets. Our analysis shows that all the types in the graph are effective for inter-sentence relation extraction.

Ameya Godbole, Aman Dalmia, Sunil Kumar Sahu

Determining whether two given questions are semantically similar is a fairly challenging task given the different structures and forms that the questions can take. In this paper, we use Gated Recurrent Units(GRU) in combination with other highly used machine learning algorithms like Random Forest, Adaboost and SVM for the similarity prediction task on a dataset released by Quora, consisting of about 400k labeled question pairs. We got the best result by using the Siamese adaptation of a Bidirectional GRU with a Random Forest classifier, which landed us among the top 24% in the competition Quora Question Pairs hosted on Kaggle.

Kushal Chawla, Sunil Kumar Sahu, Ashish Anand

Learning algorithms for natural language processing (NLP) tasks traditionally rely on manually defined relevant contextual features. On the other hand, neural network models using an only distributional representation of words have been successfully applied for several NLP tasks. Such models learn features automatically and avoid explicit feature engineering. Across several domains, neural models become a natural choice specifically when limited characteristics of data are known. However, this flexibility comes at the cost of interpretability. In this paper, we define three different methods to investigate ability of bi-directional recurrent neural networks (RNNs) in capturing contextual features. In particular, we analyze RNNs for sequence tagging tasks. We perform a comprehensive analysis on general as well as biomedical domain datasets. Our experiments focus on important contextual words as features, which can easily be extended to analyze various other feature types. We also investigate positional effects of context words and show how the developed methods can be used for error analysis.

Sunil Kumar Sahu, Ashish Anand

Lack of sufficient labeled data often limits the applicability of advanced machine learning algorithms to real life problems. However efficient use of Transfer Learning (TL) has been shown to be very useful across domains. TL utilizes valuable knowledge learned in one task (source task), where sufficient data is available, to the task of interest (target task). In biomedical and clinical domain, it is quite common that lack of sufficient training data do not allow to fully exploit machine learning models. In this work, we present two unified recurrent neural models leading to three transfer learning frameworks for relation classification tasks. We systematically investigate effectiveness of the proposed frameworks in transferring the knowledge under multiple aspects related to source and target tasks, such as, similarity or relatedness between source and target tasks, and size of training data for source task. Our empirical results show that the proposed frameworks in general improve the model performance, however these improvements do depend on aspects related to source and target tasks. This dependence then finally determine the choice of a particular TL framework.

Sunil Kumar Sahu, Ashish Anand

Simultaneous administration of multiple drugs can have synergistic or antagonistic effects as one drug can affect activities of other drugs. Synergistic effects lead to improved therapeutic outcomes, whereas, antagonistic effects can be life-threatening, may lead to increased healthcare cost, or may even cause death. Thus identification of unknown drug-drug interaction (DDI) is an important concern for efficient and effective healthcare. Although multiple resources for DDI exist, they are often unable to keep pace with rich amount of information available in fast growing biomedical texts. Most existing methods model DDI extraction from text as a classification problem and mainly rely on handcrafted features. Some of these features further depend on domain specific tools. Recently neural network models using latent features have been shown to give similar or better performance than the other existing models dependent on handcrafted features. In this paper, we present three models namely, {\it B-LSTM}, {\it AB-LSTM} and {\it Joint AB-LSTM} based on long short-term memory (LSTM) network. All three models utilize word and position embedding as latent features and thus do not rely on explicit feature engineering. Further use of bidirectional long short-term memory (Bi-LSTM) networks allow implicit feature extraction from the whole sentence. The two models, {\it AB-LSTM} and {\it Joint AB-LSTM} also use attentive pooling in the output of Bi-LSTM layer to assign weights to features. Our experimental results on the SemEval-2013 DDI extraction dataset show that the {\it Joint AB-LSTM} model outperforms all the existing methods, including those relying on handcrafted features. The other two proposed LSTM models also perform competitively with state-of-the-art methods.