AI Lab Lossfunk Enables LLMs to Generate Text in Tulu Without Prior Training

Lossfunk, an AI research lab founded by Wingify cofounder Paras Chopra, has developed a breakthrough method that enables large language models to generate text in Tulu, a coastal Karnataka language with around two million speakers, without any prior training in the language.

The approach uses a five-layer prompt of approximately 2,800 tokens that incorporates Tulu grammar rules, a list of forbidden Kannada words, and a self-verification checklist. This structured method achieved nearly 85% grammatical accuracy while reducing Kannada contamination from 80% to just 5%. The technique showed strong results across multiple models, including Gemini 2.0 Flash, GPT-4o, and Llama 3.1 70B.

The development addresses a critical challenge for low-resource languages in India that have minimal online presence and training data. Lossfunk's research suggests that prompt engineering alone may be sufficient to enable LLMs to reason in languages they were never trained on, potentially serving as a template for bringing other underrepresented Indian languages into AI systems without expensive data collection or specialized model training.