Large Language Models (LLMs) have revolutionized how we interact with and process information, but this revolution hasn't reached everyone equally. LLMs typically struggle with low-resource languages—languages with less available training data. This digital language divide limits access to cutting-edge AI technologies for many communities around the world. However, new research offers a promising solution: a more efficient way to continually pre-train these massive models, making it feasible to adapt them to under-resourced languages without breaking the bank. The challenge lies in the sheer size of LLMs and the computational resources required for training. Continual pre-training (CPT), which involves further training an existing model on new data, is a powerful technique for adapting LLMs to new domains or languages. However, for low-resource languages, collecting and processing the massive datasets needed for CPT is often prohibitively expensive. Researchers are tackling this challenge by developing clever algorithms to optimize the CPT process. Instead of blindly using all available data, they've created a system that intelligently selects the most informative sentences from a larger corpus. This 'quality over quantity' approach drastically reduces the amount of data needed, making CPT far more efficient. The algorithm identifies 'important' words within the corpus based on their contextual significance and the popularity of their subword components. By prioritizing sentences rich in these key linguistic elements, the model can learn more effectively from a smaller dataset. Further enhancing this approach is a novel method for expanding the LLM's vocabulary. This targeted vocabulary augmentation focuses on adding tokens that are both contextually relevant and representative of common linguistic structures in the low-resource language. By improving the model's understanding of crucial words and phrases, this technique leads to further performance gains, especially for languages whose scripts are not well-represented in existing LLM vocabularies. Experiments with nine Indic languages, covering diverse scripts and varying levels of resource availability, have shown promising results. Using a benchmark dataset of generation tasks, the researchers demonstrated significant improvements in LLM performance after applying their efficient CPT methods. The results highlight the potential of this approach to bridge the digital language divide and bring the power of LLMs to more communities worldwide. Interestingly, the research also revealed that adding more data and vocabulary isn't always better. The key lies in selecting the *right* data and vocabulary to maximize the impact of CPT. While this research focuses on Indic languages, the underlying principles and algorithms have broader implications. This targeted approach to continual pre-training could be adapted to other low-resource languages, opening doors for more inclusive and accessible AI technologies. The future of LLMs hinges on making them truly multilingual, and this research offers a vital stepping stone towards that goal.