Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -214,10 +214,23 @@ Two caveats stated up front, because they decide how the figure may be read.
|
|
| 214 |
|
| 215 |
- **No benchmark evaluation.** Ruled out on time grounds; published numbers cover all 40
|
| 216 |
checkpoints, so nothing was lost.
|
| 217 |
-
- `qmd` / `coordinate_gap` / `coord_fraction` are recorded NaN
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 221 |
|
| 222 |
## 5. How to reproduce
|
| 223 |
|
|
@@ -361,31 +374,33 @@ One row per (operator, task combination, property): the with-AIM value, its base
|
|
| 361 |
|
| 362 |
F8 property columns, and the reason for any that are empty. NaN-by-construction columns are kept rather than dropped.
|
| 363 |
|
| 364 |
-
| column | family | in_F8 | populated | reason_if_not
|
| 365 |
-
|:--------------------------|:---------------|:--------|:------------|:---------------------------------------------------------------------|
|
| 366 |
-
| weight_cosine | weight space | True | False | sweep incomplete
|
| 367 |
-
| qmd_raw | weight space | True | False | sweep incomplete
|
| 368 |
-
| coord_fraction | weight space | True | False | NaN by construction: coordinate_gap / qmd_raw with a zero numerator.
|
| 369 |
-
| subspace_overlap | weight space | True | False | sweep incomplete
|
| 370 |
-
| spectral_overcounting | weight space | True | False | sweep incomplete
|
| 371 |
-
| geo_cka_mean | representation | True | False | sweep incomplete
|
| 372 |
-
| geo_procrustes_mean | representation | True | False | sweep incomplete
|
| 373 |
-
| geo_subspace_overlap_mean | representation | True | False | sweep incomplete
|
| 374 |
-
| geo_svcca_mean | representation | True | False | sweep incomplete
|
| 375 |
-
| geo_cka_late_minus_early | representation | True | False | sweep incomplete
|
| 376 |
-
| ret_identity_p_at_1 | retrieval | True | False | sweep incomplete
|
| 377 |
-
| ret_procrustes_p_at_1 | retrieval | True | False | sweep incomplete
|
| 378 |
-
| ret_ridge_p_at_1 | retrieval | True | False | sweep incomplete
|
| 379 |
-
| ret_gain_over_identity | retrieval | True | False | sweep incomplete
|
| 380 |
-
| grad_cosine | gradient | True | False | sweep incomplete
|
| 381 |
-
| grad_l2 | gradient | True | False | sweep incomplete
|
| 382 |
-
| grad_norm_ratio | gradient | True | False | sweep incomplete
|
| 383 |
-
| grad_cosine_layer_min | gradient | True | False | sweep incomplete
|
| 384 |
-
| beh_js | behaviour | True | False | sweep incomplete
|
| 385 |
-
| beh_logit_cosine | behaviour | True | False | sweep incomplete
|
| 386 |
-
| beh_topk_overlap | behaviour | True | False | sweep incomplete
|
| 387 |
-
| beh_rank_corr | behaviour | True | False | sweep incomplete
|
| 388 |
-
| beh_entropy_gap | behaviour | True | False | sweep incomplete
|
|
|
|
|
|
|
| 389 |
|
| 390 |
|
| 391 |
<!-- AIM-TABLES:END -->
|
|
|
|
| 214 |
|
| 215 |
- **No benchmark evaluation.** Ruled out on time grounds; published numbers cover all 40
|
| 216 |
checkpoints, so nothing was lost.
|
| 217 |
+
- `qmd` / `coordinate_gap` / `coord_fraction` are recorded NaN, and kept as columns with the
|
| 218 |
+
reason attached rather than dropped. Every checkpoint here is `W_base + Δ`, so no permutation
|
| 219 |
+
symmetry was ever broken between a merged model and its base: the residual-basis map is the
|
| 220 |
+
identity by construction and the coordinate component is zero a priori. Measuring it would
|
| 221 |
+
measure nothing. (Same reasoning as the MergeBench sweep.) `qmd_raw` **is** measured, exactly,
|
| 222 |
+
from the norms the weight pass already accumulates.
|
| 223 |
+
- **The extended 512-sentence probe (`geoX_*`, `retX_*`) was dropped.** It is the robustness
|
| 224 |
+
check on the n/d problem — 32 mean-pooled observations against d = 5120 — not the figure, which
|
| 225 |
+
uses the canonical 32-sentence probe so its columns mean the same thing as the Beetle F8's. On
|
| 226 |
+
this box (load average ~330, three brain-preprocessing jobs and a second merge sweep sharing the
|
| 227 |
+
CPU) the 41-layer Procrustes/SVCCA geometry on 512x5120 was taking longer than every other part
|
| 228 |
+
of a pair combined, and would have pushed the sweep past the time the user has. The n/d caveat
|
| 229 |
+
therefore stands unquantified here; it is stated rather than measured.
|
| 230 |
+
- **AIM's ω is not swept.** Only ω = 0.4 checkpoints are published, and building others would
|
| 231 |
+
mean running the merges ourselves. Section 2 gives the closed form, so the ω-dependence of the
|
| 232 |
+
weight-space intervention is analytic; its effect on benchmarks is not measurable from public
|
| 233 |
+
artefacts.
|
| 234 |
|
| 235 |
## 5. How to reproduce
|
| 236 |
|
|
|
|
| 374 |
|
| 375 |
F8 property columns, and the reason for any that are empty. NaN-by-construction columns are kept rather than dropped.
|
| 376 |
|
| 377 |
+
| column | family | in_F8 | populated | reason_if_not |
|
| 378 |
+
|:--------------------------|:---------------|:--------|:------------|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
| 379 |
+
| weight_cosine | weight space | True | False | sweep incomplete |
|
| 380 |
+
| qmd_raw | weight space | True | False | sweep incomplete |
|
| 381 |
+
| coord_fraction | weight space | True | False | NaN by construction: coordinate_gap / qmd_raw with a zero numerator. |
|
| 382 |
+
| subspace_overlap | weight space | True | False | sweep incomplete |
|
| 383 |
+
| spectral_overcounting | weight space | True | False | sweep incomplete |
|
| 384 |
+
| geo_cka_mean | representation | True | False | sweep incomplete |
|
| 385 |
+
| geo_procrustes_mean | representation | True | False | sweep incomplete |
|
| 386 |
+
| geo_subspace_overlap_mean | representation | True | False | sweep incomplete |
|
| 387 |
+
| geo_svcca_mean | representation | True | False | sweep incomplete |
|
| 388 |
+
| geo_cka_late_minus_early | representation | True | False | sweep incomplete |
|
| 389 |
+
| ret_identity_p_at_1 | retrieval | True | False | sweep incomplete |
|
| 390 |
+
| ret_procrustes_p_at_1 | retrieval | True | False | sweep incomplete |
|
| 391 |
+
| ret_ridge_p_at_1 | retrieval | True | False | sweep incomplete |
|
| 392 |
+
| ret_gain_over_identity | retrieval | True | False | sweep incomplete |
|
| 393 |
+
| grad_cosine | gradient | True | False | sweep incomplete |
|
| 394 |
+
| grad_l2 | gradient | True | False | sweep incomplete |
|
| 395 |
+
| grad_norm_ratio | gradient | True | False | sweep incomplete |
|
| 396 |
+
| grad_cosine_layer_min | gradient | True | False | sweep incomplete |
|
| 397 |
+
| beh_js | behaviour | True | False | sweep incomplete |
|
| 398 |
+
| beh_logit_cosine | behaviour | True | False | sweep incomplete |
|
| 399 |
+
| beh_topk_overlap | behaviour | True | False | sweep incomplete |
|
| 400 |
+
| beh_rank_corr | behaviour | True | False | sweep incomplete |
|
| 401 |
+
| beh_entropy_gap | behaviour | True | False | sweep incomplete |
|
| 402 |
+
| geoX_* | representation | False | False | extended 512-sentence probe not run: it is the robustness check on the n/d problem, not the figure, and the 41-layer Procrustes/SVCCA geometry at 512x5120 dominated the sweep's cost on a contended box. |
|
| 403 |
+
| retX_* | retrieval | False | False | extended 512-sentence probe not run: it is the robustness check on the n/d problem, not the figure, and the 41-layer Procrustes/SVCCA geometry at 512x5120 dominated the sweep's cost on a contended box. |
|
| 404 |
|
| 405 |
|
| 406 |
<!-- AIM-TABLES:END -->
|