Submitted: 15 February (link) 

In this paper, we present TituLLMs, the first large pretrained Bangla LLMs, available in 1b and 3b parameter sizes. Due to computational constraints during both training and inference, we focused on smaller models. To train TituLLMs, we collected a pretraining dataset of approximately ~37 billion tokens. We extended the Llama-3.2 tokenizer to incorporate language- and culture-specific knowledge, which also enables faster training and inference. There was a lack of benchmarking datasets to benchmark LLMs for Bangla. To address this gap, we developed five benchmarking datasets. We benchmarked various LLMs, including TituLLMs, and demonstrated that TituLLMs outperforms its initial multilingual versions. However, this is not always the case, highlighting the complexities of language adaptation. Our work lays the groundwork for adapting existing multilingual open models to other low-resource languages. To facilitate broader adoption and further research, we have made the TituLLMs models and benchmarking datasets publicly available. (this https URL)

Published: 25 August, Heliyon (link) 

Human Pose Estimation (HPE) is a crucial step towards understanding people in images and videos. HPE provides geometric and motion information of the human body, which has been applied to a wide range of applications (e.g., human-computer interaction, motion analysis, augmented reality, virtual reality, healthcare, etc.). An extremely useful task of this kind is the 2D pose estimation of bedridden patients from infrared (IR) images. Here, the IR imaging modality is preferred due to privacy concerns and the need for monitoring both uncovered and covered patients at different levels of illumination. The major drawback of this research problem is the unavailability of covered examples, which are very costly to collect and time-consuming to label. In this work, a deep learning-based framework was developed for human sleeping pose estimation on covered images using only the uncovered training images. In the training scheme, two different image augmentation techniques, a statistical approach as well as a GAN-based approach, were explored for domain adaptation, where the statistical approach performed better. The accuracy of the model trained on the statistically augmented dataset was improved by 124 % as compared with the model trained on non-augmented images. To handle the scarcity of training infrared images, a transfer learning strategy was used by pre-training the model on an RGB pose estimation dataset, resulting in a further increment in accuracy of 4 %. Semi-supervised learning techniques, with a novel pose discriminator model in the loop, were adopted to utilize the unannotated training data, resulting in a further 3 % increase in accuracy. Thus, significant improvement has been shown in the case of 2D pose estimation from infrared images, with a comparatively small amount of annotated data and a large amount of unannotated data by using the proposed training pipeline powered by heavy augmentation.