mirror of
https://github.com/wassname/weight-steering.git
synced 2026-09-10 12:44:18 +08:00
Addresses three concerns from docs/review/v6_hypothesis_review.md: 1. R_w split into oproj/downproj + Frobenius-balanced combined. 2. dW_left_basis_ceiling as the true weight oracle. 3. axis_kind tag (write/read/mixed/ceiling). Single-seed result: chars_clusters and attn_min_taskdiff are top-5 by both R_act and R_w_combined. Write-family bases (write/mlp_write/global_write) all have R_w_combined ~ 1.0 (random null) -- natural weight-side bases fail the weight-axis test. Multi-seed deferred to v7b.
7 lines
1.0 KiB
Plaintext
7 lines
1.0 KiB
Plaintext
2026-04-26T19:50:18.413132+0800 | INFO | __main__:<module>:95 - loaded Qwen/Qwen3-0.6B | layers=28 | d_model=1024 | LoRA tensors=98 | W_PATH=out/sycophancy/lora/w.pt
|
|
2026-04-26T19:50:18.414027+0800 | INFO | __main__:<module>:308 - capturing B-side label and A-side activations
|
|
2026-04-26T19:50:29.638005+0800 | INFO | __main__:<module>:332 - captured label=(28, 16, 1024) | clean=(28, 16, 1024) | up=(28, 16, 1024) | attn_tokens=(28, 16, 59, 1024)
|
|
2026-04-26T19:53:43.500018+0800 | INFO | __main__:<module>:591 - built 42 A-side candidates + ceiling
|
|
2026-04-26T19:53:43.500192+0800 | INFO | __main__:lora_weight_tensors:624 - lora_weight_tensors layer=0 dropped: [('model.layers.0.self_attn.o_proj.weight', 'missing-from-LoRA'), ('model.layers.0.mlp.down_proj.weight', 'missing-from-LoRA')]
|
|
2026-04-26T19:54:19.211549+0800 | INFO | __main__:<module>:873 - weight ceiling (dW_left_basis): combined=16.390 oproj=22.521 downproj=10.258 SHOULD: combined ~ d_model/PCS = 128 (or close); oproj/downproj near same. ELSE per-tensor split or null normalization is wrong.
|