jevbooks

← All projects

love-arena-v2

shentonyan/love-arena-v2

Local LLM judges for love-language rewrites, testing Jev-style decisions.

Other49%Runner-up: Evaluation & BenchmarkingInspection

What it is

Tests whether three local Ollama models can judge hostile-to-love-language rewrites on PoL criteria, using fully crossed judging, self-preference correction, and calibration. Validation labels are Claude-generated, not human.

How it uses Jev

This project does not call Jev. It reproduces Jev's Choice/Score/Noul pattern in jev_local.py using Ollama logprobs: the judge emits a single option token, read from top_logprobs, with forward/reversed scales and negated Noul statements to expose position bias.

Primitives:choicescorenoul

Technique worth stealing

Order-reversed and negated re-asks on logprobs to expose position bias, with coverage flagging out-of-vocabulary probability mass.

Try it

python run.py validate --run flowtest --tuned --set holdout; python run.py analyze --run flowtest --tuned --quick

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-centricno0.11
  • Shows a System One patternunclear0.41
  • Handles uncertaintyno0.03
  • Measuredno0.10
  • Runnableno0.07
  • Worth recommendingno0.03
  • Model replicaunclear0.47
  • Problem scopescore on a 0–2 scale0.57
  • About Jevyes0.89

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 22 Sept 2026.