Junaid Hamid Bhat

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.

Sowmya Rasipuram, Junaid Hamid Bhat, Anutosh Maitra

As we exceed upon the procedures for modelling the different aspects of behaviour, expression recognition has become a key field of research in Human Computer Interactions. Expression recognition in the wild is a very interesting problem and is challenging as it involves detailed feature extraction and heavy computation. This paper presents the methodologies and techniques used in our contribution to recognize different expressions i.e., neutral, anger, disgust, fear, happiness, sadness, surprise in ABAW competition on Aff-Wild2 database. Aff-Wild2 database consists of videos in the wild labelled for seven different expressions at frame level. We used a bi-modal approach by fusing audio and visual features and train a sequence-to-sequence model that is based on Gated Recurrent Units (GRU) and Long Short Term Memory (LSTM) network. We show experimental results on validation data. The overall accuracy of the proposed approach is 41.5 \%, which is better than the competition baseline of 37\%.

Sowmya Rasipuram, Junaid Hamid Bhat, Anutosh Maitra

Continuous affect prediction in the wild is a very interesting problem and is challenging as continuous prediction involves heavy computation. This paper presents the methodologies and techniques used in our contribution to predict continuous emotion dimensions i.e., valence and arousal in ABAW competition on Aff-Wild2 database. Aff-Wild2 database consists of videos in the wild labelled for valence and arousal at frame level. Our proposed methodology uses fusion of both audio and video features (multi-modal) extracted using state-of-the-art methods. These audio-video features are used to train a sequence-to-sequence model that is based on Gated Recurrent Units (GRU). We show promising results on validation data with simple architecture. The overall valence and arousal of the proposed approach is 0.22 and 0.34, which is better than the competition baseline of 0.14 and 0.24 respectively.