jevbooks

← All projects

jev-korean-benchmark

mahlernim/jev-korean-benchmark · Homepage

Reproducible early-access check of Jev on Korean and medical text.

Evaluation & Benchmarking99%Runner-up: Classification & Routing

What it is

A small frozen study that asks 100 questions per cell from four public test sets (Belebele, PAWS-X, MedQA, KorMedMCQA) in Korean and English, recording every Jev response. It is a sample check, not a benchmark.

How it uses Jev

Jev answers multiple-choice questions: it returns a probability for every option at once plus the highest-scoring option, with no reasoning text to parse. The study records those answers and compares them across languages, tasks, and against gpt-5.6-luna.

Primitives:choice

Technique worth stealing

Ask Jev the same question in both languages to isolate language effects; reorder inputs to expose answer inconsistencies.

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.83
  • Shows a System One patternno0.20
  • Handles uncertaintyno0.04
  • Measuredyes0.97
  • Runnableno0.09
  • Worth recommendingno0.24
  • Model replicano0.05
  • Problem scopescore on a 0–2 scale0.94
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.