Releases: DaoyuanLi2816/pairjudge
Releases · DaoyuanLi2816/pairjudge
pairjudge 0.1.1
Docs-only patch release — first release published via PyPI trusted publishing.
- README (shipped in the sdist, shown on PyPI) now includes the measured position-bias study (29.2% verdict flip rate on order swap; swap debiasing improves log-loss 1.0496 → 1.0462), the token-budget packing diagram, the two-phase distillation flowchart, and the gold-medal certificate.
- New example:
examples/position_bias_experiment.py— trains a judge end to end through the public API on one consumer GPU and measures position bias. - No code changes since 0.1.0.
Install: pip install pairjudge