Reverse-engineered Jev-like model (github.com)

A new reverse-engineered Jev-like model has been released, developed to efficiently choose from a dynamic list of text options in a single pass, similar to TypeSafe's proprietary model. This open-source starter model utilizes a unique scoring mechanism where each option is evaluated based on attention weights applied to context tokens, resulting in a probability distribution over choices. The project includes examples from popular games like Doom and chess, demonstrating its applicability in gaming scenarios. Users can train the model on synthetic data, leveraging either a simple byte encoder or an optional pretrained Hugging Face encoder, which enhances performance without retraining the encoder’s weights. This release is significant for the AI/ML community as it presents a viable alternative to the closed Jev model, emphasizing the potential of lightweight models that can make decisions rapidly without the need for extensive computational resources. With reported accuracy reaching around 98% on synthetic menus and 26% on real-world next-click data, this model showcases the effectiveness of one-pass scoring, which is substantially faster than traditional decoding methods. By making advanced AI techniques more accessible, this release encourages further experimentation and innovation in the field of natural language processing and decision-making systems.