Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms

Jeff is a 0.8 billion decision model, trained locally in about 2 hours on a single RTX PRO 6000 GPU. It can make fast, well-calibrated judgments between options in response to a user-provided situation and list of options. The model can handle multiple choice, yes/no, or scoring tasks, generating a calibrated probability for each option in a single forward pass.

Developed as an independent project, Jeff uses the same request format as Jev but is not affiliated with TypeSafe, the creators of Jev. Models are developed from the open-source AutoJev recipe, with training code available on GitHub under the MIT license. The 0.8B and 2B models both undergo full-weight fine-tuning with a one-epoch batch size of 256, using cross-entropy over option letters and fitted temperature for calibration.

Jeff performs comparably to Jev and sometimes even surpasses it on smaller, zero-shot tasks. In a series of three games, the model demonstrated its ability to make decisions in under 50 milliseconds per move. Despite its size, Jeff provides a practical solution for on-site decision-making without relying on cloud GPUs or closed-model outputs.