ManPla commited on
Commit
3b3c102
·
verified ·
1 Parent(s): d1390ae

Archiva análisis retirados en v5 y desactiva el dataset viewer

Browse files
README.md CHANGED
@@ -1,28 +1,6 @@
1
  ---
2
  license: apache-2.0
3
- configs:
4
- - config_name: gauge_flip
5
- data_files: logs/fase_G/gauge_flip.csv
6
- - config_name: computado_estatico
7
- data_files: logs/fase_0/computado_estatico.csv
8
- - config_name: dissociation_D_by_layer
9
- data_files: tables/dissociation_D/by_layer.csv
10
- - config_name: decision_rota
11
- data_files: logs/decision_rota/decision_rota.csv
12
- - config_name: decision_rota_lm
13
- data_files: logs/decision_rota/decision_rota_lm.csv
14
- - config_name: poda_criterio
15
- data_files: logs/poda_criterio/poda_criterio.csv
16
- - config_name: orbita_gauge
17
- data_files: logs/orbita_gauge/orbita_gauge.csv
18
- - config_name: certificado_simultaneo
19
- data_files: logs/orbita_gauge/certificado_simultaneo.csv
20
- - config_name: ablation_A
21
- data_files: tables/ablation_A.csv
22
- - config_name: vitb_run
23
- data_files: logs/vitb_clean/attnA_*/run.csv
24
- - config_name: vitl_run
25
- data_files: logs/vitl_clean/attnA_*/run.csv
26
  ---
27
 
28
  # angular-separation-vit — results
@@ -32,6 +10,11 @@ dos podas distintas: la dirección dominante de $W_O$ bajo intervención
32
  libre, suave y dura"* ("Same function, different pruning: the
33
  dominant direction of $W_O$ under free, soft, and hard intervention").
34
 
 
 
 
 
 
35
  Code: <https://github.com/mmunozpl/angular-separation-vit>
36
  Reproduction checkpoints: <https://huggingface.co/ManPla/angular-separation-vit-checkpoints>
37
  Paper (both languages): [`paper/angular_separation_paper.pdf`](paper/angular_separation_paper.pdf)
@@ -39,41 +22,34 @@ Paper (both languages): [`paper/angular_separation_paper.pdf`](paper/angular_sep
39
  [`paper/separacion_angular_paper.pdf`](paper/separacion_angular_paper.pdf)
40
  (Spanish original).
41
 
42
- Each CSV group above is declared as its own dataset **config**
43
- (dropdown in the viewer): the underlying files have different
44
- schemas from each other, and a single merged view previously failed
45
- or showed spurious nulls across incompatible columns. Files not
46
- listed as a config (per-run auxiliary CSVs, `archive/`) are not part
47
- of any config and are only reachable via the Files tab, by design.
48
-
49
  ## Table/figure → file map
50
 
51
- | Paper item | File(s) | Viewer config |
52
- |---|---|---|
53
- | tab:gauge (gauge-flip invariance, H2) | `logs/fase_G/gauge_flip.csv` | `gauge_flip` |
54
- | tab:residuo (computed/static residue) | `logs/fase_0/computado_estatico.csv` | `computado_estatico` |
55
- | tab:inversion, fig:inversion (Q·K relocation) | `tables/dissociation_D/by_layer.csv` | `dissociation_D_by_layer` |
56
- | tab:arch (cross-architecture contrast) | `logs/fase_G/gauge_flip.csv`, `logs/fase_0/computado_estatico.csv`, `tables/dissociation_D/by_layer.csv` | `gauge_flip`, `computado_estatico`, `dissociation_D_by_layer` |
57
- | tab:divattn (angular regularizer ablation, ViT-B) | `logs/vitb_clean/attnA_{base,blanda,dura}_seed{42..46}/run.csv` | `vitb_run` |
58
- | ViT-L cost/inertia readings (§6.2) | `logs/vitl_clean/attnA_{base,blanda,dura}_seed{42..44}/run.csv` | `vitl_run` |
59
- | tab:decision (decision instability under gauge, vision) | `logs/decision_rota/decision_rota.csv` | `decision_rota` |
60
- | tab:decision (decision instability under gauge, language, Pythia-410M) | `logs/decision_rota/decision_rota_lm.csv` | `decision_rota_lm` |
61
- | tab:poda (pruning by similarity vs. random) | `logs/poda_criterio/poda_criterio.csv` | `poda_criterio` |
62
- | app:orbita (certified gauge orbit) | `logs/orbita_gauge/orbita_gauge.csv`, `logs/orbita_gauge/certificado_simultaneo.csv` | `orbita_gauge`, `certificado_simultaneo` |
63
- | `artifacts/tables/ablation_A.csv` | raw per-arch ablation table (last-epoch extraction; the paper's prose numbers use the ce>0 window instead, see `sec:setup` "Métricas") | `ablation_A` |
64
 
65
  Per-run auxiliary CSVs (`layerwise.csv`, `entropy.csv`,
66
  `head_norms.csv` under each `attnA_*` directory) are diagnostic detail
67
- behind `run.csv`, not cited individually, and not part of any config.
68
 
69
  ## `archive/`
70
 
71
- `archive/codes.csv`: spherical-code validation (kissing-number lower
72
- bounds) and prototype-detection artifacts from analyses retired
73
- before the current paper (v5.3 surgery dropped the Riesz-energy/
74
- spherical-codes apparatus and the earlier detection contribution
75
- entirely). Kept for provenance, not cited by any table or figure
76
- above, and not part of any viewer config.
77
 
78
  ## License
79
 
 
1
  ---
2
  license: apache-2.0
3
+ viewer: false
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  ---
5
 
6
  # angular-separation-vit — results
 
10
  libre, suave y dura"* ("Same function, different pruning: the
11
  dominant direction of $W_O$ under free, soft, and hard intervention").
12
 
13
+ The dataset viewer is disabled on purpose: this repository is a
14
+ collection of heterogeneous CSV logs (one schema per experiment),
15
+ meant for direct download. See the table/figure map below for which
16
+ file backs which result.
17
+
18
  Code: <https://github.com/mmunozpl/angular-separation-vit>
19
  Reproduction checkpoints: <https://huggingface.co/ManPla/angular-separation-vit-checkpoints>
20
  Paper (both languages): [`paper/angular_separation_paper.pdf`](paper/angular_separation_paper.pdf)
 
22
  [`paper/separacion_angular_paper.pdf`](paper/separacion_angular_paper.pdf)
23
  (Spanish original).
24
 
 
 
 
 
 
 
 
25
  ## Table/figure → file map
26
 
27
+ | Paper item | File(s) |
28
+ |---|---|
29
+ | tab:gauge (gauge-flip invariance, H2) | `logs/fase_G/gauge_flip.csv` |
30
+ | tab:residuo (computed/static residue) | `logs/fase_0/computado_estatico.csv` |
31
+ | tab:inversion, fig:inversion (Q·K relocation) | `tables/dissociation_D/by_layer.csv` |
32
+ | tab:arch (cross-architecture contrast) | `logs/fase_G/gauge_flip.csv`, `logs/fase_0/computado_estatico.csv`, `tables/dissociation_D/by_layer.csv` |
33
+ | tab:divattn (angular regularizer ablation, ViT-B) | `logs/vitb_clean/attnA_{base,blanda,dura}_seed{42..46}/run.csv` |
34
+ | ViT-L cost/inertia readings (§6.2) | `logs/vitl_clean/attnA_{base,blanda,dura}_seed{42..44}/run.csv` |
35
+ | tab:decision (decision instability under gauge, vision) | `logs/decision_rota/decision_rota.csv` |
36
+ | tab:decision (decision instability under gauge, language, Pythia-410M) | `logs/decision_rota/decision_rota_lm.csv` |
37
+ | tab:poda (pruning by similarity vs. random) | `logs/poda_criterio/poda_criterio.csv` |
38
+ | app:orbita (certified gauge orbit) | `logs/orbita_gauge/orbita_gauge.csv`, `logs/orbita_gauge/certificado_simultaneo.csv` |
39
+ | `artifacts/tables/ablation_A.csv` | raw per-arch ablation table (last-epoch extraction; the paper's prose numbers use the ce>0 window instead, see `sec:setup` "Métricas") |
40
 
41
  Per-run auxiliary CSVs (`layerwise.csv`, `entropy.csv`,
42
  `head_norms.csv` under each `attnA_*` directory) are diagnostic detail
43
+ behind `run.csv`, not cited individually.
44
 
45
  ## `archive/`
46
 
47
+ `archive/logs/codes.csv`: spherical-code validation (kissing-number
48
+ lower bounds) and prototype-detection artifacts from analyses
49
+ retired before the current paper (v5.3 surgery dropped the
50
+ Riesz-energy/spherical-codes apparatus and the earlier detection
51
+ contribution entirely). Kept for provenance, not cited by any table
52
+ or figure above. See `archive/README.md`.
53
 
54
  ## License
55
 
archive/README.md ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ Analyses retired during the v5 revision of the paper (kissing number
2
+ validation, Riesz energy apparatus, exploratory prototypes). Kept for
3
+ provenance; not referenced by any table or figure of the published
4
+ paper.
archive/{codes.csv → logs/codes.csv} RENAMED
File without changes