wassname 1650aa9191 add prompt_nll (free coherence proxy) + eval summary line ...

Per-token NLL over the scoring text is free since we already compute
full-sequence logits in guided_rollout / guided_rollout_batch (just
gather instead of slicing [:, -1]). Higher NLL = model less coherent
on this prompt under whatever steering is attached.

eval.py logs pmass/ppl/nll aggregate at end of guided eval so the
next run shows degradation at a glance instead of buried tqdm noise.
analyse() now exposes raw_nll and info.prompt_nll_mean.

2026-05-05 06:34:24 +08:00

..

__init__.py

API improvements: rename clifford->classic, default load_vignettes to all, add dual-axis docs, and update HF upload script

2026-05-03 07:01:28 +08:00

core.py

add prompt_nll (free coherence proxy) + eval summary line

2026-05-05 06:34:24 +08:00

data.py

feat: multi-label moral foundation ratings with z-scored frame averaging and human calibration

2026-05-03 12:48:14 +08:00

eval.py

add prompt_nll (free coherence proxy) + eval summary line

2026-05-05 06:34:24 +08:00

guided.py

add prompt_nll (free coherence proxy) + eval summary line

2026-05-05 06:34:24 +08:00