The Kaggle Master Abhinand Balachandran has launched “Tamil-Llama,” an Indic LLM engineered specifically to elevate the Tamil language domain. This AI model is built on top of Meta’s Llama 2.

Check out the GitHub Repository here.

Tamil-Llama is meticulously crafted, integrating additional Tamil tokens and harnessing the LoRA methodology for streamlined and effective training.

Sudalai Rajkumar (SRK), the Kaggle Grandmaster posted on LinkedIn about the model, and congratulated Balachandran for the achievement.

This model boasts variants with 7 billion and 13 billion parameters, signifying a significant stride forward in AI for Tamil and potentially establishing itself as the most advanced open-source LLM tailored for an Indian language to date.

The model offers four distinct iterations: Tamil LLaMA 7B, 13B, 7B Instruct, and 14B Instruct, catering to various complexities and requirements.

The research paper explains throughout the training phase, the model’s vocabulary has expanded to encompass 16,000 Tamil tokens, supplementing the original 32,000 tokens for enhanced linguistic inclusivity.

Datasets utilised in the fine-tuning phase are readily accessible within the repository, fostering transparency and collaboration in the AI community.

The project was built within a span of two months. Balachandran explained how he balanced the challenges of managing GPU expenses and navigating the intricate technicalities of constructing a state-of-the-art language model; this journey stands as a testament to Balachandran’s commitment. 

With a vision aimed at propelling Indian languages to the forefront of AI, Balachandran envisions Tamil-LLaMA as more than just a technological breakthrough.