Add nynxz Destripe and DeBand Strong to edit anatomy
Browse files- research/README.md +1 -1
research/README.md
CHANGED
|
@@ -48,7 +48,7 @@ D = Δ · G_num · (G_flat + μ·G_gen + ε·I)⁻¹ ε = ρ · tr(G_flat +
|
|
| 48 |
|
| 49 |
## Edit anatomy
|
| 50 |
|
| 51 |
-
`edit_anatomy.py` (+ `edit_anatomy_c.py` for the horizontal-detail measure) computes, for each fix and layer, the weight-change size relative to a full revert and the output-change energy tr(D·G·Dᵀ) on solid-colour vs general activations; `build_anatomy.py` draws it. Measure selectivity on the centred Grams: on raw Grams one per-token colour/brightness/modulation direction holds about half the energy and makes any edit look large.
|
| 52 |
|
| 53 |
## Training (negative result, kept for reproduction)
|
| 54 |
|
|
|
|
| 48 |
|
| 49 |
## Edit anatomy
|
| 50 |
|
| 51 |
+
`edit_anatomy.py` (+ `edit_anatomy_c.py` for the horizontal-detail measure) computes, for each fix and layer, the weight-change size relative to a full revert and the output-change energy tr(D·G·Dᵀ) on solid-colour vs general activations; `build_anatomy.py` draws it. `edit_anatomy_lora.py <tag> <lora>` adds any LoRA (lora_up/down or .diff keys; split q/k/v mapped into fused qkv) as another arm. Measure selectivity on the centred Grams: on raw Grams one per-token colour/brightness/modulation direction holds about half the energy and makes any edit look large.
|
| 52 |
|
| 53 |
## Training (negative result, kept for reproduction)
|
| 54 |
|