๐ Dev.to
A 3B model that beats a 7B: failure-driven orchestration on uncontaminated knowledge
How a fully sovereign QA stack โ local Wikipedia index, distilled Qwen2.5-3B reader, and crutches built only from measured failures โ went from 33% to 52% on facts no LLM can have memorized, and matched a zero-shot 7B more than twice its size.
TL;DR
Configuration
Post-cutoff-150 accuracy
3B naked (no retrieval)
0.0% [0โ2.5]
7B naked
0.0% (by construction โ see below)
3B + sovereign chain (start, Sep 29)
33.3%
3B + sovereign cha...


