wassnameandClaudypoo 205e96bf6a eval: add macro informedness (Youden's J) headline metric ...

Chance-corrected, argmax-only companion to mean_nll: moves when the answer
flips, not when confidence shifts. 0 = base-rate guessing, so it exposes
majority-class models that top1_acc flatters. Same flip-informedness family
as steering-lite's surgical informedness, anchored on the human argmax here.
README also points at the paired training set moral_stories_foundations.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

2026-06-18 19:47:20 +08:00

..

data

fix: package vignette data with tinymfv

2026-06-13 22:50:56 +08:00

__init__.py

refactor

2026-05-08 15:15:14 +08:00

data.py

fix: package vignette data with tinymfv

2026-06-13 22:50:56 +08:00

eval.py

eval: add macro informedness (Youden's J) headline metric

2026-06-18 19:47:20 +08:00

guided.py

guided: reuse Phase 1 KV cache for Phase 2 + drop nll_prompt

2026-05-19 04:12:20 +00:00