A Jev of All Trades: Build and Train a Jev-style Decision Model From Scratch

A Jev-style decision model was built to classify pathology images from the PathMNIST dataset by comparing an image to a set of text descriptions. This model, which uses a small adapter on a pre-existing vision-language model, was compared against the Qwen-VL large language model and traditional classifiers. The Jev-inspired model outperformed the Qwen model (67% vs 43% accuracy) and demonstrated flexibility by handling new tissue labels without retraining. However, a standard linear classifier achieved higher accuracy (88%) on the familiar, pre-trained tissue categories, highlighting a trade-off between flexibility and specialized performance.