Building KRXYOriginally published 8 min read
Building KRXY — Part 5: Measuring Answers Before Choosing a Model
KRXY compares Qwen2.5 1.5B and 3B on a Raspberry Pi, keeps its default after a frozen hold-out, builds a model library and fixes a 256-token chat cutoff.
By mihalykrich

Canonical source
https://krxy.hashnode.dev/building-krxy-measuring-answers-model-evaluations-part-5
Hashnode remains the canonical copy unless an administrator changes that preference for locally rendered articles.
Read on Hashnode