Reverse-engineered Jev-like model
Researchers have developed an open-source model that replicates the functionality of Jev, a commercial TypeSafe model designed to select the best option from a list of text choices in a single pass rather than generating responses word-by-word. The model uses an attention-based architecture that converts each option into a query vector, assigns attention weights to context tokens, and outputs probabilities for each option via a shared scoring mechanism. The repository includes demonstrations on Doom and chess tasks, along with tools for training, evaluation, and inference on custom datasets.