Today, the Institute for Computer Science, AI, and Technology (INSAIT) officially announced the launch of a new version of Bulgaria’s first large language model, BgGPT. Bulgaria is the first Central and Eastern European country to adopt a LLM in its native language.
BgGPT is not the only focus project of INSAIT, but it is a strategic one for the country, aiming to enhance the acceptance of AI in various institutions, particularly in education. Prof. Galin Tsokov, the Minister of Education and Science, emphasized that the plan for digitalizing education would be further enhanced with the integration of the locally developed generative AI.
Martin Vechev added that INSAIT is planning training sessions with university teachers to help them use the right prompts when using it in their practice.
Vechev also explained that the new model has been further developed to minimize hallucinations and outperforms previous models created by INSAIT. “The more we use it, the better it becomes. It is crucial to rely on our own model because we train it with local data that is not accessible to overseas companies. This makes it more accurate in the information it provides,” he noted.
BgGPT models come in 3 sizes (2.6B, 9B, 27B), which are built on top of Google’s Gemma-2 base family, but with extensions, including new research. According to their insights, despite being significantly smaller, BgGPT surpasses Meta Llama 70B and Alibaba Qwen 72B for Bulgarian, while retaining English skills.
For chatting, according to GPT-4o used as a judge, BgGPT 27B surpasses both OpenAI’s free model GPT-4o-mini and Anthropic’s Haiku and is comparable to GPT-4o (paid) and Sonnet, Anthropic’s large model. The evaluation was done using 1000’s of real-world conversations, spanning around 100 different topics.
“We observe similar results with Anthropic’s Claude Haiku and Sonnet paid models,” INSAIT shared in their official announcement.
Image Credit: INSAIT; (from left to right): Martin Vechev, Dimitar Glavchev, Prof. Galin Tsokov, Prof. Georgi Vachev, Borislav Petrov
Pioneering AI language model
The announcement of the new Bulgaria’s large language model, BgGPT, took place at Sofia University, led by Martin Vechev, Scientific Director of INSAIT, Dimitar Glavchev, Bulgarian Prime Minister, Prof. Galin Tsokov, the Minister of Education and Science, Prof. Georgi Vachev, the rector of Sofia University “St. Kliment Ohridski”, and Borislav Petrov, Executive Director of INSAIT.
INSAIT is a collaboration between Sofia University “St. Kliment Ohridski”, ETH Zurich, and EPFL Lausanne. Backed by $100M from the Bulgarian government and additional funding from SiteGround, Amazon Web Services (AWS), Google, Bulgarian entrepreneurs, and DeepMind, alongside business angels Vassil Terziev, Bogomil Balkansky, Ivan Osmak, Atanas Simeonov, Nedelcho Spasov, Stanimir Vassilev, Svetozar Georgiev, Hristo Hristov, Lubo Minchev, and Ivo Evgeniev.
Bulgaria is the first CEE country with a separate AI language model, developed by a public organization. Vechev shared that INSAIT is currently working with other countries in the region in developing LLMs in their native languages.
BgGPT is not a corporate property and can be used freely by public and private organizations. It is going to be available from 23rd of November at bggpt.ai.
Earlier this year, in March, was the first launch of BgGPT.
0 comments