Tiny Jev-like decision head built atop Qwen2.5-0.5B that can be trained and run locally on Apple Silicon MacBooks.
Attaches a parallel decision head to an open 0.5B model to answer discrete questions directly from token activations.
Provides a small local decision-model architecture and probes for packed-versus-separate questions, option order and calibration. Hosted teacher calls for data collection remain a separate network and cost dependency.
The pinned README and hosted-teacher adapter were read. This site did not train the student or reproduce its calibration figures. Local inference capability does not mean every data-generation step is offline or free.
Explore a small local decision-head architecture and its training/evaluation workflow.
Material analysed：
The README describes each question seeing the document but not sibling questions. Packed-versus-separate equivalence should be checked independently from task accuracy.
Inspect the supporting source ↗kev/jev.py wraps a Vercel Gateway Jev call for teacher data and token accounting. Local student inference does not mean every training-data operation is offline or free.
Inspect the supporting source ↗Useful for learning and adapting a small decision model when local deployment matters. Compare a trained checkpoint on your own held-out questions before treating API compatibility as quality parity.
Inspect the README's packed/separate and permutation probes, then read kev/jev.py for hosted credentials and token usage before running data-generation steps.