v8: rank-honest pct_oracle metric (energy_frac / oracle@r_eff in [0,1])

Browse Source

Replaces v7's post-hoc 'pct_w_oracle = R_w / R_w_ceiling' (a ratio of two
concentration ratios) with a per-row pct_oracle: candidate's energy_frac
divided by the optimal rank-r_eff subspace's energy_frac on the same
target. Rank-honest: chars_clusters (r_eff=7) is graded against rank-7
oracle, not rank-8. Activation oracle = PCA of L2-normalized hs_diff_B
(matches existing energy_frac_act formula).

Result: every non-oracle candidate lands at pct_oracle in [0.02, 0.11] on
both axes. Best joint = WNR_union_TaskDiff at 0.089 (rank 16; all others
rank 8). chars_clusters and layer_clean_resid_pca tied at ~0.085. This is
a clean negative result: LoRA's task-specific delta is far from any of
our hand-built linear primitives' spans.

...

This commit is contained in:

wassname

2026-04-26 20:59:27 +08:00

parent 651ad132d3

commit f4039dd2ee

2 changed files with 2669 additions and 0 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

nbs/hypothesis_sweep_v8.ipynb

+1388

File diff suppressed because it is too large Load Diff

nbs/hypothesis_sweep_v8.py

+1281

File diff suppressed because it is too large Load Diff