kev: Jev-like model built on Qwen2.5-0.5B (github.com)

The recent launch of kev, a Jev-like decision model built on the Qwen2.5-0.5B architecture, marks a significant advancement in handling decision-making tasks within the AI/ML community. This innovative system utilizes a LoRA adapter and a unique readout head to enable the model to process and respond to multiple typed questions directly from a single document input, all in one efficient forward pass. The implementation of a block-causal masking technique ensures that each question is processed independently, enhancing accuracy and integrity in the decision-making process. kev's design is notable for its reliance on calibrated probabilities rather than generating textual responses, a shift that could influence how models are trained for decision-making tasks. Trained using cross-entropy against labeled outcomes, kev has demonstrated impressive performance metrics, including held-out expected calibration errors as low as 0.031 after parameter scaling. Its ease of integration is bolstered by a drop-in API compatible with TypeSafe's ecosystem, allowing for smooth implementation on local environments. With a training duration of approximately 1 hour and 45 minutes on an Apple M5 and a fast response time for queries, kev shows promising potential for applications that require rapid and accurate decision support, ranging from customer service to data analysis.