Sinhala AI Breakthrough: IEEE Access Publishes New LLM

A Sri Lankan research team has made a significant breakthrough in Sinhala artificial intelligence. They've developed a working Sinhala language model using just two GPUs, a stark contrast to the billions spent by global tech giants. What's interesting is how well it performs. In a blind evaluation, this new model scored 4.5 out of 5 for Sinhala prompts, while the Meta Llama 3.1 model scored only 1 out of 5. The team also reduced the model's perplexity, a measure of language understanding, by nearly 90 percent. This means the model can hold natural conversations, answer questions, and follow instructions coherently in Sinhala. The research has been accepted for publication in IEEE Access, a prestigious peer-reviewed journal. This is notable because the journal has a strict binary review policy, accepting very few low-resource-language AI papers. The bottom line is this breakthrough offers a sovereign Sinhala LLM. This could enable locally hosted AI assistants for government services, educational tools, healthcare information, and customer service, providing critical control over data privacy and national security for Sri Lanka.