The new VaidhLlama 3B parameter model addresses critical limitations in general LLMs regarding Sanskrit-heavy Ayurveda logic. Developed by a team including Vivek Patel, it achieves 41.91% accuracy on the BhashaBench-Ayur benchmark, outperforming baselines like Llama-3.2-3B and Gemma-2-27B. The architecture uses NVIDIA NeMo Curator for high-density data curation and vLLM for synthetic data generation.

A key innovation is 'Unmasked Instruction Tuning,' which trains on full sequences to capture complex question syntax. This specialization improves clinical domains by 100% but introduces a 'Specialist's Curse' trade-off, reducing general knowledge. The project is integrated with the BharatGen ecosystem and aims for future scaling at IIT Bombay.