lev: a Qwen 4B decision model with dual readouts

A Qwen3.5-4B LoRA adapter answers multiple typed questions without generating answer tokens. Its runtime combines label-token readout with a learned candidate head, option-order averaging and saved calibration profiles.

Notes

The author reports macro accuracy 68.9% over all 3,880 items across 13 pinned S1Bench subsets, with mean ECE 0.115; released weights use the TypeSafe /v1/systemone protocol.

source-backed — Public repository, docs or live artifact. About this label

Destinations

Also linked

• lev model card and adapter ↗

Filed under

• #local
• #training
• #calibration
• #evaluation

In collections

• Benchmarks & independent tests →
• Local & replica models →