[NEWS] CHOONOK Company Unveils Veterinary-Specific LLM, “VetJarvis-4B-Instruct”
Performance Validated Using a Benchmark Based on Japan’s National Veterinary Licensing Examination
Choonok Company has unveiled VetJarvis-4B-Instruct, a generative large language model (LLM) specialized for the veterinary domain, marking a new milestone in the development of veterinary AI for research and education.
The release demonstrates that a generative AI model built around veterinary knowledge can achieve meaningful performance in professional knowledge assessment environments.
VetJarvis-4B-Instruct was released to support the expansion of the veterinary research and education ecosystem. Its model weights are available through the Hugging Face platform. The model is intended primarily for research and educational use, with clear usage guidelines stating that it is not designed for clinical decision-making.
Built on Veterinary-Specific Knowledge
VetJarvis-4B-Instruct was trained on datasets developed from veterinary academic literature, clinical guidelines, and expert knowledge accumulated through real-world clinical practice.
The training data spans a wide range of clinical disciplines, including internal medicine, surgery, diagnostic imaging, and oncology. Personally identifiable and other identifiable information was removed through a de-identification process before the data was used.
The model was developed through collaboration between veterinarians with clinical experience and technical development teams. Validation was also conducted to better reflect the terminology and decision-making processes used in real-world veterinary practice.
In particular, veterinary knowledge was incorporated from the pre-training stage, rather than relying solely on post-training adaptation. This approach was designed to improve the model’s understanding of the veterinary domain at a more fundamental level.
Objective Evaluation of a Domain-Specific AI Model
VetJarvis-4B-Instruct was evaluated using questions derived from previous editions of Japan’s National Veterinary Licensing Examination, providing a benchmark for assessing its veterinary knowledge and problem-solving capabilities.
The model achieved competitive accuracy compared with general-purpose models of a similar scale, demonstrating the potential of domain-specific AI architectures to perform effectively in specialized professional fields.
The model is also designed to avoid overly confident responses when available evidence is uncertain or insufficient. In such cases, it indicates the need for additional review rather than presenting unsupported conclusions as definitive.
This design reflects Choonok Company’s principle that AI should support expert judgment rather than replace it.
The release of VetJarvis-4B-Instruct represents an important technical step toward making generative AI more practical and reliable for veterinary research and education.
Building VetBench and Developing the Next Generation of Veterinary AI Models
Following the release of VetJarvis-4B-Instruct, Choonok Company is preparing VetBench, a benchmark framework designed to evaluate the performance of veterinary AI models.
VetBench aims to provide a standardized environment in which different models can be evaluated under the same criteria, helping improve the reliability, comparability, and reproducibility of veterinary AI research.
Choonok Company is also developing a larger next-generation model, with plans for release in the second half of this year. The company will continue expanding its technology for use across a broader range of clinical specialties and research environments.
뉴스 읽어보기
Comments