V-REX: Efficient Specialist VLM Training for Veterinary X-Rays

TL;DR - V-REX is a specialist vision-language model trained from scratch to generate diagnostic reports for veterinary radiographs. It challenges the assumption that domain expertise requires fine-tuning large foundation models, reporting better task performance with substantially fewer parameters, data, and compute.

• Redesigns the full VLM pipeline, including tokenization, generative pre-training, grounding, and inference.
• Introduces training strategies intended to improve data utilization, efficiency, and downstream performance.
• Uses only veterinary radiology data rather than relying on external foundation-model training data.
• Reports significant gains over open foundation models, though the provided abstract gives no quantitative results.