mirror of
https://github.com/wassname/weight-steering.git
synced 2026-06-27 15:18:02 +08:00
wassname
21 lines
833 B
Python
21 lines
833 B
Python
"""SVD + AntiPaSTO subspace alignment for the diff vector w.
|
|
|
|
Per layer:
|
|
W = U_out @ diag(S) @ U_in.T # SVD of pretrained weight
|
|
w_layer = θ+_layer - θ-_layer # the steering diff
|
|
|
|
Subspaces tested (see docs/AntiPaSTO_concepts/):
|
|
- SVD top-k : column span of U_out[:, :k] (or U_in[:, :k])
|
|
- Suppressed : PCA of layer-to-layer magnitude drops on a probe set
|
|
- Write-not-read : col_span(W_o, W_down) ∩ orth(row_span(W_q,W_k,W_v,W_up_{L+1}))
|
|
- Weak-readout : bottom-1% of unembedding SVD
|
|
- Stenographic : task_diff ∩ suppressed
|
|
|
|
Metric per layer:
|
|
energy_ratio = ||proj_subspace(w_layer)||² / ||w_layer||²
|
|
null = same projection applied to a random rank-r matrix scaled to ||w_layer||
|
|
alignment = energy_ratio / null with bootstrap CI
|
|
|
|
A ratio > 1 (CI excluding 1) is real alignment.
|
|
"""
|