jjho-choi commited on
Commit
b03e94a
·
verified ·
1 Parent(s): bed4dd4

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. vision/dinov2-b_layer10/mp/x16/l0_128/code_metrics_step_final.json +136 -0
  2. vision/dinov2-b_layer10/mp/x16/l0_128/dict_metrics_step_final.json +545 -0
  3. vision/dinov2-b_layer10/mp/x16/l0_128/seed0/config.json +18 -0
  4. vision/dinov2-b_layer10/mp/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  5. vision/dinov2-b_layer10/mp/x16/l0_128/seed1/config.json +18 -0
  6. vision/dinov2-b_layer10/mp/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  7. vision/dinov2-b_layer10/mp/x16/l0_128/seed2/config.json +18 -0
  8. vision/dinov2-b_layer10/mp/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  9. vision/dinov2-b_layer10/mp/x16/l0_32/code_metrics_step_final.json +136 -0
  10. vision/dinov2-b_layer10/mp/x16/l0_32/dict_metrics_step_final.json +545 -0
  11. vision/dinov2-b_layer10/mp/x16/l0_32/seed0/config.json +18 -0
  12. vision/dinov2-b_layer10/mp/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  13. vision/dinov2-b_layer10/mp/x16/l0_32/seed1/config.json +18 -0
  14. vision/dinov2-b_layer10/mp/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  15. vision/dinov2-b_layer10/mp/x16/l0_32/seed2/config.json +18 -0
  16. vision/dinov2-b_layer10/mp/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  17. vision/dinov2-b_layer10/mp/x16/l0_64/code_metrics_step_final.json +136 -0
  18. vision/dinov2-b_layer10/mp/x16/l0_64/dict_metrics_step_final.json +545 -0
  19. vision/dinov2-b_layer10/mp/x16/l0_64/seed0/config.json +18 -0
  20. vision/dinov2-b_layer10/mp/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  21. vision/dinov2-b_layer10/mp/x16/l0_64/seed1/config.json +18 -0
  22. vision/dinov2-b_layer10/mp/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  23. vision/dinov2-b_layer10/mp/x16/l0_64/seed2/config.json +18 -0
  24. vision/dinov2-b_layer10/mp/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  25. vision/dinov2-b_layer10/mp/x4/l0_128/code_metrics_step_final.json +136 -0
  26. vision/dinov2-b_layer10/mp/x4/l0_128/dict_metrics_step_final.json +595 -0
  27. vision/dinov2-b_layer10/mp/x4/l0_32/code_metrics_step_final.json +136 -0
  28. vision/dinov2-b_layer10/mp/x4/l0_32/dict_metrics_step_final.json +595 -0
  29. vision/dinov2-b_layer10/mp/x4/l0_32/seed1/config.json +18 -0
  30. vision/dinov2-b_layer10/mp/x4/l0_64/code_metrics_step_final.json +136 -0
  31. vision/dinov2-b_layer10/mp/x4/l0_64/dict_metrics_step_final.json +595 -0
  32. vision/dinov2-b_layer10/mp/x4/l0_64/seed0/config.json +18 -0
  33. vision/dinov2-b_layer10/mp/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  34. vision/dinov2-b_layer10/mp/x4/l0_64/seed1/config.json +18 -0
  35. vision/dinov2-b_layer10/mp/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  36. vision/dinov2-b_layer10/mp/x4/l0_64/seed2/config.json +18 -0
  37. vision/dinov2-b_layer10/mp/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  38. vision/dinov2-b_layer10/relu/x16/l0_128/seed0/config.json +10 -0
  39. vision/dinov2-b_layer10/relu/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  40. vision/dinov2-b_layer10/relu/x16/l0_128/seed1/config.json +10 -0
  41. vision/dinov2-b_layer10/relu/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  42. vision/dinov2-b_layer10/relu/x16/l0_128/seed2/config.json +10 -0
  43. vision/dinov2-b_layer10/relu/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json +9 -0
  44. vision/dinov2-b_layer10/relu/x16/l0_32/seed0/config.json +10 -0
  45. vision/dinov2-b_layer10/relu/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  46. vision/dinov2-b_layer10/relu/x16/l0_32/seed1/config.json +10 -0
  47. vision/dinov2-b_layer10/relu/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  48. vision/dinov2-b_layer10/relu/x16/l0_32/seed2/config.json +10 -0
  49. vision/dinov2-b_layer10/relu/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json +15 -0
  50. vision/dinov2-b_layer10/relu/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json +15 -0
vision/dinov2-b_layer10/mp/x16/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 303.350046692607,
20
+ "recon_r2": 0.9945499165194519,
21
+ "l1_per_token": 235.00570939926905
22
+ },
23
+ "seed1": {
24
+ "l0": 308.26556420233464,
25
+ "recon_r2": 0.9947738428369166,
26
+ "l1_per_token": 237.94410117398962
27
+ },
28
+ "seed2": {
29
+ "l0": 303.32809208819714,
30
+ "recon_r2": 0.994642080225655,
31
+ "l1_per_token": 236.53932976929894
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.3081740635698948,
38
+ "mcc_abs": 0.308183846174609,
39
+ "act_iou": 0.06158082759211115,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9944915843011223,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.486311554204309,
44
+ "0.7": 0.47742297952236057,
45
+ "0.8": 0.4724732855685592,
46
+ "0.9": 0.499113829069671
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.48502962515759485,
50
+ "0.7": 0.47719360709953557,
51
+ "0.8": 0.4724564351388078,
52
+ "0.9": 0.49935382096494135
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.30419426695625673,
58
+ "mcc_abs": 0.30420468250052374,
59
+ "act_iou": 0.06069953939073782,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.994404892243357,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.48488945895927105,
64
+ "0.7": 0.47440591114644926,
65
+ "0.8": 0.4711649853133228,
66
+ "0.9": 0.49944654239146674
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.4840749925413885,
70
+ "0.7": 0.4744411480351126,
71
+ "0.8": 0.46963779513166715,
72
+ "0.9": 0.4997708880092602
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.3070745853473141,
78
+ "mcc_abs": 0.3070853782622768,
79
+ "act_iou": 0.06065747285174358,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.994486634330592,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.4903236837017785,
84
+ "0.7": 0.48064124130373964,
85
+ "0.8": 0.4743521280493735,
86
+ "0.9": 0.4958878012745798
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.48972885496687907,
90
+ "0.7": 0.4805835291198603,
91
+ "0.8": 0.4730622072347284,
92
+ "0.9": 0.49558014937754513
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 304.9812343277129,
106
+ "l0_std": 2.8443342885749576,
107
+ "recon_r2_mean": 0.9946552798606746,
108
+ "recon_r2_std": 0.00011254519841739937,
109
+ "l1_per_token_mean": 236.49638011418588,
110
+ "l1_per_token_std": 1.4696666493093344,
111
+ "input_r2_mean": 0.9944610369583572,
112
+ "input_r2_std": 4.868569932631574e-05,
113
+ "mcc_signed_mean": 0.30648097195782187,
114
+ "mcc_signed_std": 0.0020552318393844463,
115
+ "mcc_abs_mean": 0.30649130231246985,
116
+ "mcc_abs_std": 0.0020550256839986587,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.4871748989551195,
119
+ "0.7": 0.4774900439908498,
120
+ "0.8": 0.4726634663104185,
121
+ "0.9": 0.49814939091190585
122
+ },
123
+ "matched_mcc_signed_band": 0.48386945004207343,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.48627782422195415,
126
+ "0.7": 0.47740609475150286,
127
+ "0.8": 0.4717188125017344,
128
+ "0.9": 0.49823495278391555
129
+ },
130
+ "matched_mcc_abs_band": 0.48340942106477675,
131
+ "act_iou_mean": 0.060979279944864184,
132
+ "act_iou_std": 0.0005213799740659474,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,545 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 8607,
34
+ "0.6": 7559,
35
+ "0.7": 6051,
36
+ "0.8": 3854,
37
+ "0.9": 1228
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 2371,
41
+ "0.6": 2294,
42
+ "0.7": 2002,
43
+ "0.8": 1407,
44
+ "0.9": 519
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 367,
48
+ "0.6": 236,
49
+ "0.7": 134,
50
+ "0.8": 29,
51
+ "0.9": 10
52
+ },
53
+ "babel_mu1_signed": 0.6670389771461487,
54
+ "n_dup_abs": {
55
+ "0.5": 8643,
56
+ "0.6": 7607,
57
+ "0.7": 6081,
58
+ "0.8": 3874,
59
+ "0.9": 1228
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 2386,
63
+ "0.6": 2311,
64
+ "0.7": 2013,
65
+ "0.8": 1414,
66
+ "0.9": 517
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 372,
70
+ "0.6": 295,
71
+ "0.7": 114,
72
+ "0.8": 29,
73
+ "0.9": 10
74
+ },
75
+ "babel_mu1_abs": 0.6695666313171387
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": null,
81
+ "l0_raw": null,
82
+ "recon_r2": null,
83
+ "l1_per_token": null,
84
+ "stats_source": "mp-excluded",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 0,
89
+ "dense_rate": 0.0,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 8532,
93
+ "0.6": 7493,
94
+ "0.7": 6077,
95
+ "0.8": 3854,
96
+ "0.9": 1170
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 2322,
100
+ "0.6": 2234,
101
+ "0.7": 2022,
102
+ "0.8": 1405,
103
+ "0.9": 494
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 439,
107
+ "0.6": 325,
108
+ "0.7": 145,
109
+ "0.8": 50,
110
+ "0.9": 10
111
+ },
112
+ "babel_mu1_signed": 0.6654933094978333,
113
+ "n_dup_abs": {
114
+ "0.5": 8570,
115
+ "0.6": 7525,
116
+ "0.7": 6107,
117
+ "0.8": 3874,
118
+ "0.9": 1172
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 2326,
122
+ "0.6": 2245,
123
+ "0.7": 2036,
124
+ "0.8": 1414,
125
+ "0.9": 494
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 235,
129
+ "0.6": 173,
130
+ "0.7": 147,
131
+ "0.8": 50,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_abs": 0.667901337146759
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": null,
140
+ "l0_raw": null,
141
+ "recon_r2": null,
142
+ "l1_per_token": null,
143
+ "stats_source": "mp-excluded",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 0,
148
+ "dense_rate": 0.0,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 8456,
152
+ "0.6": 7442,
153
+ "0.7": 6006,
154
+ "0.8": 3793,
155
+ "0.9": 1148
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 2327,
159
+ "0.6": 2258,
160
+ "0.7": 1995,
161
+ "0.8": 1400,
162
+ "0.9": 485
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 355,
166
+ "0.6": 144,
167
+ "0.7": 77,
168
+ "0.8": 36,
169
+ "0.9": 10
170
+ },
171
+ "babel_mu1_signed": 0.6603737473487854,
172
+ "n_dup_abs": {
173
+ "0.5": 8512,
174
+ "0.6": 7487,
175
+ "0.7": 6050,
176
+ "0.8": 3818,
177
+ "0.9": 1152
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 2346,
181
+ "0.6": 2272,
182
+ "0.7": 2011,
183
+ "0.8": 1409,
184
+ "0.9": 485
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 338,
188
+ "0.6": 286,
189
+ "0.7": 88,
190
+ "0.8": 36,
191
+ "0.9": 10
192
+ },
193
+ "babel_mu1_abs": 0.6630961298942566
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7817156910896301,
199
+ "dcs_abs": 0.7834139466285706,
200
+ "frac_sign_flipped": 0.039388020833333336,
201
+ "alive_source": "stats",
202
+ "n_alive_a": 12288,
203
+ "n_alive_b": 12288,
204
+ "matched_cos_signed": {
205
+ "0.5": 0.7236672043800354,
206
+ "0.6": 0.7293717265129089,
207
+ "0.7": 0.7402892112731934,
208
+ "0.8": 0.7550824880599976,
209
+ "0.9": 0.7738751173019409
210
+ },
211
+ "unique_frac_signed": {
212
+ "0.5": 0.4935709635416667,
213
+ "0.6": 0.57177734375,
214
+ "0.7": 0.6702473958333334,
215
+ "0.8": 0.80078125,
216
+ "0.9": 0.9436442057291666
217
+ },
218
+ "dedup_dcs_signed": {
219
+ "0.5": 0.3571811193493583,
220
+ "0.6": 0.4170382283919025,
221
+ "0.7": 0.4961769160193702,
222
+ "0.8": 0.6046558986417949,
223
+ "0.9": 0.7302627703999557
224
+ },
225
+ "matched_frac_signed": {
226
+ "0.5": 0.3589680989583333,
227
+ "0.6": 0.4026692708333333,
228
+ "0.7": 0.4476725260416667,
229
+ "0.8": 0.482421875,
230
+ "0.9": 0.443359375
231
+ },
232
+ "matched_cos_abs": {
233
+ "0.5": 0.7263489365577698,
234
+ "0.6": 0.732623279094696,
235
+ "0.7": 0.7427051663398743,
236
+ "0.8": 0.7570679783821106,
237
+ "0.9": 0.7756255269050598
238
+ },
239
+ "unique_frac_abs": {
240
+ "0.5": 0.4913330078125,
241
+ "0.6": 0.5696614583333334,
242
+ "0.7": 0.6688232421875,
243
+ "0.8": 0.7998046875,
244
+ "0.9": 0.9434814453125
245
+ },
246
+ "dedup_dcs_abs": {
247
+ "0.5": 0.35687920772033976,
248
+ "0.6": 0.41734724557803327,
249
+ "0.7": 0.4967384773408412,
250
+ "0.8": 0.6055065178661607,
251
+ "0.9": 0.7317882931456552
252
+ },
253
+ "matched_frac_abs": {
254
+ "0.5": 0.3590494791666667,
255
+ "0.6": 0.4029947916666667,
256
+ "0.7": 0.447998046875,
257
+ "0.8": 0.4825032552083333,
258
+ "0.9": 0.443359375
259
+ }
260
+ },
261
+ "seed0\u2194seed2": {
262
+ "dcs_signed": 0.7770745754241943,
263
+ "dcs_abs": 0.7787559628486633,
264
+ "frac_sign_flipped": 0.043701171875,
265
+ "alive_source": "stats",
266
+ "n_alive_a": 12288,
267
+ "n_alive_b": 12288,
268
+ "matched_cos_signed": {
269
+ "0.5": 0.7226144671440125,
270
+ "0.6": 0.7275781631469727,
271
+ "0.7": 0.7358648180961609,
272
+ "0.8": 0.7517416477203369,
273
+ "0.9": 0.7691615223884583
274
+ },
275
+ "unique_frac_signed": {
276
+ "0.5": 0.4968668619791667,
277
+ "0.6": 0.5748291015625,
278
+ "0.7": 0.6720377604166666,
279
+ "0.8": 0.8030598958333334,
280
+ "0.9": 0.9441731770833334
281
+ },
282
+ "dedup_dcs_signed": {
283
+ "0.5": 0.3590431827105931,
284
+ "0.6": 0.41823310183826834,
285
+ "0.7": 0.49452894432276173,
286
+ "0.8": 0.6036935693118721,
287
+ "0.9": 0.726221678283764
288
+ },
289
+ "matched_frac_signed": {
290
+ "0.5": 0.3600260416666667,
291
+ "0.6": 0.4007161458333333,
292
+ "0.7": 0.4436848958333333,
293
+ "0.8": 0.4747721354166667,
294
+ "0.9": 0.4336751302083333
295
+ },
296
+ "matched_cos_abs": {
297
+ "0.5": 0.7251178026199341,
298
+ "0.6": 0.7300339341163635,
299
+ "0.7": 0.737865686416626,
300
+ "0.8": 0.7535355687141418,
301
+ "0.9": 0.7708365321159363
302
+ },
303
+ "unique_frac_abs": {
304
+ "0.5": 0.4945068359375,
305
+ "0.6": 0.5723063151041666,
306
+ "0.7": 0.6701253255208334,
307
+ "0.8": 0.8018798828125,
308
+ "0.9": 0.9439290364583334
309
+ },
310
+ "dedup_dcs_abs": {
311
+ "0.5": 0.35857571025553625,
312
+ "0.6": 0.417803030735134,
313
+ "0.7": 0.4944624833005946,
314
+ "0.8": 0.6042450135355466,
315
+ "0.9": 0.7276149850270789
316
+ },
317
+ "matched_frac_abs": {
318
+ "0.5": 0.3597005208333333,
319
+ "0.6": 0.4002278645833333,
320
+ "0.7": 0.443115234375,
321
+ "0.8": 0.475830078125,
322
+ "0.9": 0.4337565104166667
323
+ }
324
+ },
325
+ "seed1\u2194seed2": {
326
+ "dcs_signed": 0.7794721722602844,
327
+ "dcs_abs": 0.781230628490448,
328
+ "frac_sign_flipped": 0.042073567708333336,
329
+ "alive_source": "stats",
330
+ "n_alive_a": 12288,
331
+ "n_alive_b": 12288,
332
+ "matched_cos_signed": {
333
+ "0.5": 0.722134530544281,
334
+ "0.6": 0.7282166481018066,
335
+ "0.7": 0.7373058795928955,
336
+ "0.8": 0.7540035247802734,
337
+ "0.9": 0.7709962725639343
338
+ },
339
+ "unique_frac_signed": {
340
+ "0.5": 0.4979248046875,
341
+ "0.6": 0.5750732421875,
342
+ "0.7": 0.6717936197916666,
343
+ "0.8": 0.802978515625,
344
+ "0.9": 0.9455159505208334
345
+ },
346
+ "dedup_dcs_signed": {
347
+ "0.5": 0.3595686950793606,
348
+ "0.6": 0.4187779088388197,
349
+ "0.7": 0.49531738574538997,
350
+ "0.8": 0.6054486311040819,
351
+ "0.9": 0.7289892735013079
352
+ },
353
+ "matched_frac_signed": {
354
+ "0.5": 0.3595377604166667,
355
+ "0.6": 0.3990885416666667,
356
+ "0.7": 0.4423828125,
357
+ "0.8": 0.47998046875,
358
+ "0.9": 0.4520670572916667
359
+ },
360
+ "matched_cos_abs": {
361
+ "0.5": 0.7249367237091064,
362
+ "0.6": 0.7299293875694275,
363
+ "0.7": 0.7390979528427124,
364
+ "0.8": 0.7557072639465332,
365
+ "0.9": 0.7726919054985046
366
+ },
367
+ "unique_frac_abs": {
368
+ "0.5": 0.4950358072916667,
369
+ "0.6": 0.5729573567708334,
370
+ "0.7": 0.6700032552083334,
371
+ "0.8": 0.8018798828125,
372
+ "0.9": 0.9452718098958334
373
+ },
374
+ "dedup_dcs_abs": {
375
+ "0.5": 0.3588696362567134,
376
+ "0.6": 0.41821841253113234,
377
+ "0.7": 0.4951980343224326,
378
+ "0.8": 0.605986452254001,
379
+ "0.9": 0.7304038760024317
380
+ },
381
+ "matched_frac_abs": {
382
+ "0.5": 0.359130859375,
383
+ "0.6": 0.3982747395833333,
384
+ "0.7": 0.442138671875,
385
+ "0.8": 0.48046875,
386
+ "0.9": 0.452392578125
387
+ }
388
+ }
389
+ },
390
+ "summary": {
391
+ "n_seeds": 3,
392
+ "n_pairs": 3,
393
+ "report_band": [
394
+ 0.6,
395
+ 0.7,
396
+ 0.8,
397
+ 0.9
398
+ ],
399
+ "n_dead_mean": 0,
400
+ "n_dead_std": 0.0,
401
+ "n_dense_mean": 0,
402
+ "n_dense_std": 0.0,
403
+ "l0_mean": null,
404
+ "l0_std": null,
405
+ "recon_r2_mean": null,
406
+ "recon_r2_std": null,
407
+ "l1_per_token_mean": null,
408
+ "l1_per_token_std": null,
409
+ "babel_mu1_signed_mean": 0.6643020113309225,
410
+ "babel_mu1_signed_std": 0.0034886552645365565,
411
+ "n_dup_signed_per_threshold": {
412
+ "0.5": 8531.666666666666,
413
+ "0.6": 7498.0,
414
+ "0.7": 6044.666666666667,
415
+ "0.8": 3833.6666666666665,
416
+ "0.9": 1182.0
417
+ },
418
+ "n_dup_signed_band": 4639.583333333334,
419
+ "n_dup_signed_band_std": 38.65418261111381,
420
+ "n_dup_clusters_signed_per_threshold": {
421
+ "0.5": 2340.0,
422
+ "0.6": 2262.0,
423
+ "0.7": 2006.3333333333333,
424
+ "0.8": 1404.0,
425
+ "0.9": 499.3333333333333
426
+ },
427
+ "n_dup_clusters_signed_band": 1542.9166666666665,
428
+ "n_dup_clusters_signed_band_std": 11.10273990208423,
429
+ "largest_dup_signed_per_threshold": {
430
+ "0.5": 387.0,
431
+ "0.6": 235.0,
432
+ "0.7": 118.66666666666667,
433
+ "0.8": 38.333333333333336,
434
+ "0.9": 10.0
435
+ },
436
+ "largest_dup_signed_band": 100.5,
437
+ "largest_dup_signed_band_std": 32.90991491936739,
438
+ "dcs_signed_mean": 0.779420812924703,
439
+ "dcs_signed_std": 0.0023209840566022644,
440
+ "dedup_dcs_signed_per_threshold": {
441
+ "0.5": 0.35859766571310403,
442
+ "0.6": 0.41801641302299686,
443
+ "0.7": 0.49534108202917393,
444
+ "0.8": 0.6045993663525829,
445
+ "0.9": 0.7284912407283425
446
+ },
447
+ "dedup_dcs_signed_band": 0.561612025533274,
448
+ "dedup_dcs_signed_band_std": 0.000817928943294489,
449
+ "matched_cos_signed_per_threshold": {
450
+ "0.5": 0.722805400689443,
451
+ "0.6": 0.7283888459205627,
452
+ "0.7": 0.7378199696540833,
453
+ "0.8": 0.7536092201868693,
454
+ "0.9": 0.7713443040847778
455
+ },
456
+ "matched_cos_signed_band": 0.7477905849615732,
457
+ "matched_cos_signed_band_std": 0.001789422150281412,
458
+ "unique_frac_signed_per_threshold": {
459
+ "0.5": 0.49612087673611116,
460
+ "0.6": 0.5738932291666666,
461
+ "0.7": 0.6713595920138888,
462
+ "0.8": 0.8022732204861112,
463
+ "0.9": 0.9444444444444445
464
+ },
465
+ "unique_frac_signed_band": 0.7479926215277778,
466
+ "unique_frac_signed_band_std": 0.0012055337535602764,
467
+ "matched_frac_signed_per_threshold": {
468
+ "0.5": 0.3595106336805556,
469
+ "0.6": 0.40082465277777773,
470
+ "0.7": 0.444580078125,
471
+ "0.8": 0.47905815972222227,
472
+ "0.9": 0.4430338541666667
473
+ },
474
+ "matched_frac_signed_band": 0.4418741861979167,
475
+ "matched_frac_signed_band_std": 0.003188141718741717,
476
+ "babel_mu1_abs_mean": 0.6668546994527181,
477
+ "babel_mu1_abs_std": 0.0033598266343399697,
478
+ "n_dup_abs_per_threshold": {
479
+ "0.5": 8575.0,
480
+ "0.6": 7539.666666666667,
481
+ "0.7": 6079.333333333333,
482
+ "0.8": 3855.3333333333335,
483
+ "0.9": 1184.0
484
+ },
485
+ "n_dup_abs_band": 4664.583333333333,
486
+ "n_dup_abs_band_std": 35.63033585771166,
487
+ "n_dup_clusters_abs_per_threshold": {
488
+ "0.5": 2352.6666666666665,
489
+ "0.6": 2276.0,
490
+ "0.7": 2020.0,
491
+ "0.8": 1412.3333333333333,
492
+ "0.9": 498.6666666666667
493
+ },
494
+ "n_dup_clusters_abs_band": 1551.75,
495
+ "n_dup_clusters_abs_band_std": 10.5,
496
+ "largest_dup_abs_per_threshold": {
497
+ "0.5": 315.0,
498
+ "0.6": 251.33333333333334,
499
+ "0.7": 116.33333333333333,
500
+ "0.8": 38.333333333333336,
501
+ "0.9": 10.0
502
+ },
503
+ "largest_dup_abs_band": 104.0,
504
+ "largest_dup_abs_band_std": 8.54400374531753,
505
+ "dcs_abs_mean": 0.7811335126558939,
506
+ "dcs_abs_std": 0.0023305099951428116,
507
+ "dedup_dcs_abs_per_threshold": {
508
+ "0.5": 0.3581081847441965,
509
+ "0.6": 0.4177895629480999,
510
+ "0.7": 0.49546633165462284,
511
+ "0.8": 0.6052459945519028,
512
+ "0.9": 0.7299357180583886
513
+ },
514
+ "dedup_dcs_abs_band": 0.5621094018032535,
515
+ "dedup_dcs_abs_band_std": 0.0009540964043521437,
516
+ "matched_cos_abs_per_threshold": {
517
+ "0.5": 0.7254678209622701,
518
+ "0.6": 0.7308622002601624,
519
+ "0.7": 0.7398896018664042,
520
+ "0.8": 0.7554369370142618,
521
+ "0.9": 0.7730513215065002
522
+ },
523
+ "matched_cos_abs_band": 0.7498100151618321,
524
+ "matched_cos_abs_band_std": 0.0020075506852886187,
525
+ "unique_frac_abs_per_threshold": {
526
+ "0.5": 0.4936252170138889,
527
+ "0.6": 0.5716417100694445,
528
+ "0.7": 0.669650607638889,
529
+ "0.8": 0.8011881510416666,
530
+ "0.9": 0.9442274305555557
531
+ },
532
+ "unique_frac_abs_band": 0.746676974826389,
533
+ "unique_frac_abs_band_std": 0.0010942126864736965,
534
+ "matched_frac_abs_per_threshold": {
535
+ "0.5": 0.3592936197916667,
536
+ "0.6": 0.4004991319444444,
537
+ "0.7": 0.4444173177083333,
538
+ "0.8": 0.4796006944444444,
539
+ "0.9": 0.44316948784722227
540
+ },
541
+ "matched_frac_abs_band": 0.4419216579861111,
542
+ "matched_frac_abs_band_std": 0.003226171902808814,
543
+ "alive_source": "stats"
544
+ }
545
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed0/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1298,
3
+ "r2_score": 0.9648,
4
+ "l0_per_token_0": 0.0029,
5
+ "l0_per_token_0.001": 0.0029,
6
+ "l0_per_token_0.01": 0.0029,
7
+ "l0_per_token_0.05": 0.0029,
8
+ "l0_per_token_0.1": 0.0029,
9
+ "l0_per_token_0.5": 0.0029,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 96.7825
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed1/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1258,
3
+ "r2_score": 0.9659,
4
+ "l0_per_token_0": 0.0031,
5
+ "l0_per_token_0.001": 0.0031,
6
+ "l0_per_token_0.01": 0.0031,
7
+ "l0_per_token_0.05": 0.0031,
8
+ "l0_per_token_0.1": 0.0031,
9
+ "l0_per_token_0.5": 0.0031,
10
+ "l0_per_token_1.0": 0.0027,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 100.9962
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed2/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1287,
3
+ "r2_score": 0.9651,
4
+ "l0_per_token_0": 0.003,
5
+ "l0_per_token_0.001": 0.003,
6
+ "l0_per_token_0.01": 0.003,
7
+ "l0_per_token_0.05": 0.003,
8
+ "l0_per_token_0.1": 0.003,
9
+ "l0_per_token_0.5": 0.003,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 98.9773
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 303.350046692607,
20
+ "recon_r2": 0.9945499165194519,
21
+ "l1_per_token": 235.00570939926905
22
+ },
23
+ "seed1": {
24
+ "l0": 308.26556420233464,
25
+ "recon_r2": 0.9947738428369166,
26
+ "l1_per_token": 237.94410117398962
27
+ },
28
+ "seed2": {
29
+ "l0": 303.32809208819714,
30
+ "recon_r2": 0.994642080225655,
31
+ "l1_per_token": 236.53932976929894
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.3081740635698948,
38
+ "mcc_abs": 0.308183846174609,
39
+ "act_iou": 0.06158082759211115,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9944915843011223,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.486311554204309,
44
+ "0.7": 0.47742297952236057,
45
+ "0.8": 0.4724732855685592,
46
+ "0.9": 0.499113829069671
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.48502962515759485,
50
+ "0.7": 0.47719360709953557,
51
+ "0.8": 0.4724564351388078,
52
+ "0.9": 0.49935382096494135
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.30419426695625673,
58
+ "mcc_abs": 0.30420468250052374,
59
+ "act_iou": 0.06069953939073782,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.994404892243357,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.48488945895927105,
64
+ "0.7": 0.47440591114644926,
65
+ "0.8": 0.4711649853133228,
66
+ "0.9": 0.49944654239146674
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.4840749925413885,
70
+ "0.7": 0.4744411480351126,
71
+ "0.8": 0.46963779513166715,
72
+ "0.9": 0.4997708880092602
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.3070745853473141,
78
+ "mcc_abs": 0.3070853782622768,
79
+ "act_iou": 0.06065747285174358,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.994486634330592,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.4903236837017785,
84
+ "0.7": 0.48064124130373964,
85
+ "0.8": 0.4743521280493735,
86
+ "0.9": 0.4958878012745798
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.48972885496687907,
90
+ "0.7": 0.4805835291198603,
91
+ "0.8": 0.4730622072347284,
92
+ "0.9": 0.49558014937754513
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 304.9812343277129,
106
+ "l0_std": 2.8443342885749576,
107
+ "recon_r2_mean": 0.9946552798606746,
108
+ "recon_r2_std": 0.00011254519841739937,
109
+ "l1_per_token_mean": 236.49638011418588,
110
+ "l1_per_token_std": 1.4696666493093344,
111
+ "input_r2_mean": 0.9944610369583572,
112
+ "input_r2_std": 4.868569932631574e-05,
113
+ "mcc_signed_mean": 0.30648097195782187,
114
+ "mcc_signed_std": 0.0020552318393844463,
115
+ "mcc_abs_mean": 0.30649130231246985,
116
+ "mcc_abs_std": 0.0020550256839986587,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.4871748989551195,
119
+ "0.7": 0.4774900439908498,
120
+ "0.8": 0.4726634663104185,
121
+ "0.9": 0.49814939091190585
122
+ },
123
+ "matched_mcc_signed_band": 0.48386945004207343,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.48627782422195415,
126
+ "0.7": 0.47740609475150286,
127
+ "0.8": 0.4717188125017344,
128
+ "0.9": 0.49823495278391555
129
+ },
130
+ "matched_mcc_abs_band": 0.48340942106477675,
131
+ "act_iou_mean": 0.060979279944864184,
132
+ "act_iou_std": 0.0005213799740659474,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,545 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 8607,
34
+ "0.6": 7559,
35
+ "0.7": 6051,
36
+ "0.8": 3854,
37
+ "0.9": 1228
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 2371,
41
+ "0.6": 2294,
42
+ "0.7": 2002,
43
+ "0.8": 1407,
44
+ "0.9": 519
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 367,
48
+ "0.6": 236,
49
+ "0.7": 134,
50
+ "0.8": 29,
51
+ "0.9": 10
52
+ },
53
+ "babel_mu1_signed": 0.6670389771461487,
54
+ "n_dup_abs": {
55
+ "0.5": 8643,
56
+ "0.6": 7607,
57
+ "0.7": 6081,
58
+ "0.8": 3874,
59
+ "0.9": 1228
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 2386,
63
+ "0.6": 2311,
64
+ "0.7": 2013,
65
+ "0.8": 1414,
66
+ "0.9": 517
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 372,
70
+ "0.6": 295,
71
+ "0.7": 114,
72
+ "0.8": 29,
73
+ "0.9": 10
74
+ },
75
+ "babel_mu1_abs": 0.6695666313171387
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": null,
81
+ "l0_raw": null,
82
+ "recon_r2": null,
83
+ "l1_per_token": null,
84
+ "stats_source": "mp-excluded",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 0,
89
+ "dense_rate": 0.0,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 8532,
93
+ "0.6": 7493,
94
+ "0.7": 6077,
95
+ "0.8": 3854,
96
+ "0.9": 1170
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 2322,
100
+ "0.6": 2234,
101
+ "0.7": 2022,
102
+ "0.8": 1405,
103
+ "0.9": 494
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 439,
107
+ "0.6": 325,
108
+ "0.7": 145,
109
+ "0.8": 50,
110
+ "0.9": 10
111
+ },
112
+ "babel_mu1_signed": 0.6654933094978333,
113
+ "n_dup_abs": {
114
+ "0.5": 8570,
115
+ "0.6": 7525,
116
+ "0.7": 6107,
117
+ "0.8": 3874,
118
+ "0.9": 1172
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 2326,
122
+ "0.6": 2245,
123
+ "0.7": 2036,
124
+ "0.8": 1414,
125
+ "0.9": 494
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 235,
129
+ "0.6": 173,
130
+ "0.7": 147,
131
+ "0.8": 50,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_abs": 0.667901337146759
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": null,
140
+ "l0_raw": null,
141
+ "recon_r2": null,
142
+ "l1_per_token": null,
143
+ "stats_source": "mp-excluded",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 0,
148
+ "dense_rate": 0.0,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 8456,
152
+ "0.6": 7442,
153
+ "0.7": 6006,
154
+ "0.8": 3793,
155
+ "0.9": 1148
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 2327,
159
+ "0.6": 2258,
160
+ "0.7": 1995,
161
+ "0.8": 1400,
162
+ "0.9": 485
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 355,
166
+ "0.6": 144,
167
+ "0.7": 77,
168
+ "0.8": 36,
169
+ "0.9": 10
170
+ },
171
+ "babel_mu1_signed": 0.6603737473487854,
172
+ "n_dup_abs": {
173
+ "0.5": 8512,
174
+ "0.6": 7487,
175
+ "0.7": 6050,
176
+ "0.8": 3818,
177
+ "0.9": 1152
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 2346,
181
+ "0.6": 2272,
182
+ "0.7": 2011,
183
+ "0.8": 1409,
184
+ "0.9": 485
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 338,
188
+ "0.6": 286,
189
+ "0.7": 88,
190
+ "0.8": 36,
191
+ "0.9": 10
192
+ },
193
+ "babel_mu1_abs": 0.6630961298942566
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7817156910896301,
199
+ "dcs_abs": 0.7834139466285706,
200
+ "frac_sign_flipped": 0.039388020833333336,
201
+ "alive_source": "stats",
202
+ "n_alive_a": 12288,
203
+ "n_alive_b": 12288,
204
+ "matched_cos_signed": {
205
+ "0.5": 0.7236672043800354,
206
+ "0.6": 0.7293717265129089,
207
+ "0.7": 0.7402892112731934,
208
+ "0.8": 0.7550824880599976,
209
+ "0.9": 0.7738751173019409
210
+ },
211
+ "unique_frac_signed": {
212
+ "0.5": 0.4935709635416667,
213
+ "0.6": 0.57177734375,
214
+ "0.7": 0.6702473958333334,
215
+ "0.8": 0.80078125,
216
+ "0.9": 0.9436442057291666
217
+ },
218
+ "dedup_dcs_signed": {
219
+ "0.5": 0.3571811193493583,
220
+ "0.6": 0.4170382283919025,
221
+ "0.7": 0.4961769160193702,
222
+ "0.8": 0.6046558986417949,
223
+ "0.9": 0.7302627703999557
224
+ },
225
+ "matched_frac_signed": {
226
+ "0.5": 0.3589680989583333,
227
+ "0.6": 0.4026692708333333,
228
+ "0.7": 0.4476725260416667,
229
+ "0.8": 0.482421875,
230
+ "0.9": 0.443359375
231
+ },
232
+ "matched_cos_abs": {
233
+ "0.5": 0.7263489365577698,
234
+ "0.6": 0.732623279094696,
235
+ "0.7": 0.7427051663398743,
236
+ "0.8": 0.7570679783821106,
237
+ "0.9": 0.7756255269050598
238
+ },
239
+ "unique_frac_abs": {
240
+ "0.5": 0.4913330078125,
241
+ "0.6": 0.5696614583333334,
242
+ "0.7": 0.6688232421875,
243
+ "0.8": 0.7998046875,
244
+ "0.9": 0.9434814453125
245
+ },
246
+ "dedup_dcs_abs": {
247
+ "0.5": 0.35687920772033976,
248
+ "0.6": 0.41734724557803327,
249
+ "0.7": 0.4967384773408412,
250
+ "0.8": 0.6055065178661607,
251
+ "0.9": 0.7317882931456552
252
+ },
253
+ "matched_frac_abs": {
254
+ "0.5": 0.3590494791666667,
255
+ "0.6": 0.4029947916666667,
256
+ "0.7": 0.447998046875,
257
+ "0.8": 0.4825032552083333,
258
+ "0.9": 0.443359375
259
+ }
260
+ },
261
+ "seed0\u2194seed2": {
262
+ "dcs_signed": 0.7770745754241943,
263
+ "dcs_abs": 0.7787559628486633,
264
+ "frac_sign_flipped": 0.043701171875,
265
+ "alive_source": "stats",
266
+ "n_alive_a": 12288,
267
+ "n_alive_b": 12288,
268
+ "matched_cos_signed": {
269
+ "0.5": 0.7226144671440125,
270
+ "0.6": 0.7275781631469727,
271
+ "0.7": 0.7358648180961609,
272
+ "0.8": 0.7517416477203369,
273
+ "0.9": 0.7691615223884583
274
+ },
275
+ "unique_frac_signed": {
276
+ "0.5": 0.4968668619791667,
277
+ "0.6": 0.5748291015625,
278
+ "0.7": 0.6720377604166666,
279
+ "0.8": 0.8030598958333334,
280
+ "0.9": 0.9441731770833334
281
+ },
282
+ "dedup_dcs_signed": {
283
+ "0.5": 0.3590431827105931,
284
+ "0.6": 0.41823310183826834,
285
+ "0.7": 0.49452894432276173,
286
+ "0.8": 0.6036935693118721,
287
+ "0.9": 0.726221678283764
288
+ },
289
+ "matched_frac_signed": {
290
+ "0.5": 0.3600260416666667,
291
+ "0.6": 0.4007161458333333,
292
+ "0.7": 0.4436848958333333,
293
+ "0.8": 0.4747721354166667,
294
+ "0.9": 0.4336751302083333
295
+ },
296
+ "matched_cos_abs": {
297
+ "0.5": 0.7251178026199341,
298
+ "0.6": 0.7300339341163635,
299
+ "0.7": 0.737865686416626,
300
+ "0.8": 0.7535355687141418,
301
+ "0.9": 0.7708365321159363
302
+ },
303
+ "unique_frac_abs": {
304
+ "0.5": 0.4945068359375,
305
+ "0.6": 0.5723063151041666,
306
+ "0.7": 0.6701253255208334,
307
+ "0.8": 0.8018798828125,
308
+ "0.9": 0.9439290364583334
309
+ },
310
+ "dedup_dcs_abs": {
311
+ "0.5": 0.35857571025553625,
312
+ "0.6": 0.417803030735134,
313
+ "0.7": 0.4944624833005946,
314
+ "0.8": 0.6042450135355466,
315
+ "0.9": 0.7276149850270789
316
+ },
317
+ "matched_frac_abs": {
318
+ "0.5": 0.3597005208333333,
319
+ "0.6": 0.4002278645833333,
320
+ "0.7": 0.443115234375,
321
+ "0.8": 0.475830078125,
322
+ "0.9": 0.4337565104166667
323
+ }
324
+ },
325
+ "seed1\u2194seed2": {
326
+ "dcs_signed": 0.7794721722602844,
327
+ "dcs_abs": 0.781230628490448,
328
+ "frac_sign_flipped": 0.042073567708333336,
329
+ "alive_source": "stats",
330
+ "n_alive_a": 12288,
331
+ "n_alive_b": 12288,
332
+ "matched_cos_signed": {
333
+ "0.5": 0.722134530544281,
334
+ "0.6": 0.7282166481018066,
335
+ "0.7": 0.7373058795928955,
336
+ "0.8": 0.7540035247802734,
337
+ "0.9": 0.7709962725639343
338
+ },
339
+ "unique_frac_signed": {
340
+ "0.5": 0.4979248046875,
341
+ "0.6": 0.5750732421875,
342
+ "0.7": 0.6717936197916666,
343
+ "0.8": 0.802978515625,
344
+ "0.9": 0.9455159505208334
345
+ },
346
+ "dedup_dcs_signed": {
347
+ "0.5": 0.3595686950793606,
348
+ "0.6": 0.4187779088388197,
349
+ "0.7": 0.49531738574538997,
350
+ "0.8": 0.6054486311040819,
351
+ "0.9": 0.7289892735013079
352
+ },
353
+ "matched_frac_signed": {
354
+ "0.5": 0.3595377604166667,
355
+ "0.6": 0.3990885416666667,
356
+ "0.7": 0.4423828125,
357
+ "0.8": 0.47998046875,
358
+ "0.9": 0.4520670572916667
359
+ },
360
+ "matched_cos_abs": {
361
+ "0.5": 0.7249367237091064,
362
+ "0.6": 0.7299293875694275,
363
+ "0.7": 0.7390979528427124,
364
+ "0.8": 0.7557072639465332,
365
+ "0.9": 0.7726919054985046
366
+ },
367
+ "unique_frac_abs": {
368
+ "0.5": 0.4950358072916667,
369
+ "0.6": 0.5729573567708334,
370
+ "0.7": 0.6700032552083334,
371
+ "0.8": 0.8018798828125,
372
+ "0.9": 0.9452718098958334
373
+ },
374
+ "dedup_dcs_abs": {
375
+ "0.5": 0.3588696362567134,
376
+ "0.6": 0.41821841253113234,
377
+ "0.7": 0.4951980343224326,
378
+ "0.8": 0.605986452254001,
379
+ "0.9": 0.7304038760024317
380
+ },
381
+ "matched_frac_abs": {
382
+ "0.5": 0.359130859375,
383
+ "0.6": 0.3982747395833333,
384
+ "0.7": 0.442138671875,
385
+ "0.8": 0.48046875,
386
+ "0.9": 0.452392578125
387
+ }
388
+ }
389
+ },
390
+ "summary": {
391
+ "n_seeds": 3,
392
+ "n_pairs": 3,
393
+ "report_band": [
394
+ 0.6,
395
+ 0.7,
396
+ 0.8,
397
+ 0.9
398
+ ],
399
+ "n_dead_mean": 0,
400
+ "n_dead_std": 0.0,
401
+ "n_dense_mean": 0,
402
+ "n_dense_std": 0.0,
403
+ "l0_mean": null,
404
+ "l0_std": null,
405
+ "recon_r2_mean": null,
406
+ "recon_r2_std": null,
407
+ "l1_per_token_mean": null,
408
+ "l1_per_token_std": null,
409
+ "babel_mu1_signed_mean": 0.6643020113309225,
410
+ "babel_mu1_signed_std": 0.0034886552645365565,
411
+ "n_dup_signed_per_threshold": {
412
+ "0.5": 8531.666666666666,
413
+ "0.6": 7498.0,
414
+ "0.7": 6044.666666666667,
415
+ "0.8": 3833.6666666666665,
416
+ "0.9": 1182.0
417
+ },
418
+ "n_dup_signed_band": 4639.583333333334,
419
+ "n_dup_signed_band_std": 38.65418261111381,
420
+ "n_dup_clusters_signed_per_threshold": {
421
+ "0.5": 2340.0,
422
+ "0.6": 2262.0,
423
+ "0.7": 2006.3333333333333,
424
+ "0.8": 1404.0,
425
+ "0.9": 499.3333333333333
426
+ },
427
+ "n_dup_clusters_signed_band": 1542.9166666666665,
428
+ "n_dup_clusters_signed_band_std": 11.10273990208423,
429
+ "largest_dup_signed_per_threshold": {
430
+ "0.5": 387.0,
431
+ "0.6": 235.0,
432
+ "0.7": 118.66666666666667,
433
+ "0.8": 38.333333333333336,
434
+ "0.9": 10.0
435
+ },
436
+ "largest_dup_signed_band": 100.5,
437
+ "largest_dup_signed_band_std": 32.90991491936739,
438
+ "dcs_signed_mean": 0.779420812924703,
439
+ "dcs_signed_std": 0.0023209840566022644,
440
+ "dedup_dcs_signed_per_threshold": {
441
+ "0.5": 0.35859766571310403,
442
+ "0.6": 0.41801641302299686,
443
+ "0.7": 0.49534108202917393,
444
+ "0.8": 0.6045993663525829,
445
+ "0.9": 0.7284912407283425
446
+ },
447
+ "dedup_dcs_signed_band": 0.561612025533274,
448
+ "dedup_dcs_signed_band_std": 0.000817928943294489,
449
+ "matched_cos_signed_per_threshold": {
450
+ "0.5": 0.722805400689443,
451
+ "0.6": 0.7283888459205627,
452
+ "0.7": 0.7378199696540833,
453
+ "0.8": 0.7536092201868693,
454
+ "0.9": 0.7713443040847778
455
+ },
456
+ "matched_cos_signed_band": 0.7477905849615732,
457
+ "matched_cos_signed_band_std": 0.001789422150281412,
458
+ "unique_frac_signed_per_threshold": {
459
+ "0.5": 0.49612087673611116,
460
+ "0.6": 0.5738932291666666,
461
+ "0.7": 0.6713595920138888,
462
+ "0.8": 0.8022732204861112,
463
+ "0.9": 0.9444444444444445
464
+ },
465
+ "unique_frac_signed_band": 0.7479926215277778,
466
+ "unique_frac_signed_band_std": 0.0012055337535602764,
467
+ "matched_frac_signed_per_threshold": {
468
+ "0.5": 0.3595106336805556,
469
+ "0.6": 0.40082465277777773,
470
+ "0.7": 0.444580078125,
471
+ "0.8": 0.47905815972222227,
472
+ "0.9": 0.4430338541666667
473
+ },
474
+ "matched_frac_signed_band": 0.4418741861979167,
475
+ "matched_frac_signed_band_std": 0.003188141718741717,
476
+ "babel_mu1_abs_mean": 0.6668546994527181,
477
+ "babel_mu1_abs_std": 0.0033598266343399697,
478
+ "n_dup_abs_per_threshold": {
479
+ "0.5": 8575.0,
480
+ "0.6": 7539.666666666667,
481
+ "0.7": 6079.333333333333,
482
+ "0.8": 3855.3333333333335,
483
+ "0.9": 1184.0
484
+ },
485
+ "n_dup_abs_band": 4664.583333333333,
486
+ "n_dup_abs_band_std": 35.63033585771166,
487
+ "n_dup_clusters_abs_per_threshold": {
488
+ "0.5": 2352.6666666666665,
489
+ "0.6": 2276.0,
490
+ "0.7": 2020.0,
491
+ "0.8": 1412.3333333333333,
492
+ "0.9": 498.6666666666667
493
+ },
494
+ "n_dup_clusters_abs_band": 1551.75,
495
+ "n_dup_clusters_abs_band_std": 10.5,
496
+ "largest_dup_abs_per_threshold": {
497
+ "0.5": 315.0,
498
+ "0.6": 251.33333333333334,
499
+ "0.7": 116.33333333333333,
500
+ "0.8": 38.333333333333336,
501
+ "0.9": 10.0
502
+ },
503
+ "largest_dup_abs_band": 104.0,
504
+ "largest_dup_abs_band_std": 8.54400374531753,
505
+ "dcs_abs_mean": 0.7811335126558939,
506
+ "dcs_abs_std": 0.0023305099951428116,
507
+ "dedup_dcs_abs_per_threshold": {
508
+ "0.5": 0.3581081847441965,
509
+ "0.6": 0.4177895629480999,
510
+ "0.7": 0.49546633165462284,
511
+ "0.8": 0.6052459945519028,
512
+ "0.9": 0.7299357180583886
513
+ },
514
+ "dedup_dcs_abs_band": 0.5621094018032535,
515
+ "dedup_dcs_abs_band_std": 0.0009540964043521437,
516
+ "matched_cos_abs_per_threshold": {
517
+ "0.5": 0.7254678209622701,
518
+ "0.6": 0.7308622002601624,
519
+ "0.7": 0.7398896018664042,
520
+ "0.8": 0.7554369370142618,
521
+ "0.9": 0.7730513215065002
522
+ },
523
+ "matched_cos_abs_band": 0.7498100151618321,
524
+ "matched_cos_abs_band_std": 0.0020075506852886187,
525
+ "unique_frac_abs_per_threshold": {
526
+ "0.5": 0.4936252170138889,
527
+ "0.6": 0.5716417100694445,
528
+ "0.7": 0.669650607638889,
529
+ "0.8": 0.8011881510416666,
530
+ "0.9": 0.9442274305555557
531
+ },
532
+ "unique_frac_abs_band": 0.746676974826389,
533
+ "unique_frac_abs_band_std": 0.0010942126864736965,
534
+ "matched_frac_abs_per_threshold": {
535
+ "0.5": 0.3592936197916667,
536
+ "0.6": 0.4004991319444444,
537
+ "0.7": 0.4444173177083333,
538
+ "0.8": 0.4796006944444444,
539
+ "0.9": 0.44316948784722227
540
+ },
541
+ "matched_frac_abs_band": 0.4419216579861111,
542
+ "matched_frac_abs_band_std": 0.003226171902808814,
543
+ "alive_source": "stats"
544
+ }
545
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed0/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1298,
3
+ "r2_score": 0.9648,
4
+ "l0_per_token_0": 0.0029,
5
+ "l0_per_token_0.001": 0.0029,
6
+ "l0_per_token_0.01": 0.0029,
7
+ "l0_per_token_0.05": 0.0029,
8
+ "l0_per_token_0.1": 0.0029,
9
+ "l0_per_token_0.5": 0.0029,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 96.7825
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed1/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1258,
3
+ "r2_score": 0.9659,
4
+ "l0_per_token_0": 0.0031,
5
+ "l0_per_token_0.001": 0.0031,
6
+ "l0_per_token_0.01": 0.0031,
7
+ "l0_per_token_0.05": 0.0031,
8
+ "l0_per_token_0.1": 0.0031,
9
+ "l0_per_token_0.5": 0.0031,
10
+ "l0_per_token_1.0": 0.0027,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 100.9962
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed2/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1287,
3
+ "r2_score": 0.9651,
4
+ "l0_per_token_0": 0.003,
5
+ "l0_per_token_0.001": 0.003,
6
+ "l0_per_token_0.01": 0.003,
7
+ "l0_per_token_0.05": 0.003,
8
+ "l0_per_token_0.1": 0.003,
9
+ "l0_per_token_0.5": 0.003,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 98.9773
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 303.350046692607,
20
+ "recon_r2": 0.9945499165194519,
21
+ "l1_per_token": 235.00570939926905
22
+ },
23
+ "seed1": {
24
+ "l0": 308.26556420233464,
25
+ "recon_r2": 0.9947738428369166,
26
+ "l1_per_token": 237.94410117398962
27
+ },
28
+ "seed2": {
29
+ "l0": 303.32809208819714,
30
+ "recon_r2": 0.994642080225655,
31
+ "l1_per_token": 236.53932976929894
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.3081740635698948,
38
+ "mcc_abs": 0.308183846174609,
39
+ "act_iou": 0.06158082759211115,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9944915843011223,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.486311554204309,
44
+ "0.7": 0.47742297952236057,
45
+ "0.8": 0.4724732855685592,
46
+ "0.9": 0.499113829069671
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.48502962515759485,
50
+ "0.7": 0.47719360709953557,
51
+ "0.8": 0.4724564351388078,
52
+ "0.9": 0.49935382096494135
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.30419426695625673,
58
+ "mcc_abs": 0.30420468250052374,
59
+ "act_iou": 0.06069953939073782,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.994404892243357,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.48488945895927105,
64
+ "0.7": 0.47440591114644926,
65
+ "0.8": 0.4711649853133228,
66
+ "0.9": 0.49944654239146674
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.4840749925413885,
70
+ "0.7": 0.4744411480351126,
71
+ "0.8": 0.46963779513166715,
72
+ "0.9": 0.4997708880092602
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.3070745853473141,
78
+ "mcc_abs": 0.3070853782622768,
79
+ "act_iou": 0.06065747285174358,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.994486634330592,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.4903236837017785,
84
+ "0.7": 0.48064124130373964,
85
+ "0.8": 0.4743521280493735,
86
+ "0.9": 0.4958878012745798
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.48972885496687907,
90
+ "0.7": 0.4805835291198603,
91
+ "0.8": 0.4730622072347284,
92
+ "0.9": 0.49558014937754513
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 304.9812343277129,
106
+ "l0_std": 2.8443342885749576,
107
+ "recon_r2_mean": 0.9946552798606746,
108
+ "recon_r2_std": 0.00011254519841739937,
109
+ "l1_per_token_mean": 236.49638011418588,
110
+ "l1_per_token_std": 1.4696666493093344,
111
+ "input_r2_mean": 0.9944610369583572,
112
+ "input_r2_std": 4.868569932631574e-05,
113
+ "mcc_signed_mean": 0.30648097195782187,
114
+ "mcc_signed_std": 0.0020552318393844463,
115
+ "mcc_abs_mean": 0.30649130231246985,
116
+ "mcc_abs_std": 0.0020550256839986587,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.4871748989551195,
119
+ "0.7": 0.4774900439908498,
120
+ "0.8": 0.4726634663104185,
121
+ "0.9": 0.49814939091190585
122
+ },
123
+ "matched_mcc_signed_band": 0.48386945004207343,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.48627782422195415,
126
+ "0.7": 0.47740609475150286,
127
+ "0.8": 0.4717188125017344,
128
+ "0.9": 0.49823495278391555
129
+ },
130
+ "matched_mcc_abs_band": 0.48340942106477675,
131
+ "act_iou_mean": 0.060979279944864184,
132
+ "act_iou_std": 0.0005213799740659474,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/dict_metrics_step_final.json ADDED
@@ -0,0 +1,545 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 8607,
34
+ "0.6": 7559,
35
+ "0.7": 6051,
36
+ "0.8": 3854,
37
+ "0.9": 1228
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 2371,
41
+ "0.6": 2294,
42
+ "0.7": 2002,
43
+ "0.8": 1407,
44
+ "0.9": 519
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 367,
48
+ "0.6": 236,
49
+ "0.7": 134,
50
+ "0.8": 29,
51
+ "0.9": 10
52
+ },
53
+ "babel_mu1_signed": 0.6670389771461487,
54
+ "n_dup_abs": {
55
+ "0.5": 8643,
56
+ "0.6": 7607,
57
+ "0.7": 6081,
58
+ "0.8": 3874,
59
+ "0.9": 1228
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 2386,
63
+ "0.6": 2311,
64
+ "0.7": 2013,
65
+ "0.8": 1414,
66
+ "0.9": 517
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 372,
70
+ "0.6": 295,
71
+ "0.7": 114,
72
+ "0.8": 29,
73
+ "0.9": 10
74
+ },
75
+ "babel_mu1_abs": 0.6695666313171387
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": null,
81
+ "l0_raw": null,
82
+ "recon_r2": null,
83
+ "l1_per_token": null,
84
+ "stats_source": "mp-excluded",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 0,
89
+ "dense_rate": 0.0,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 8532,
93
+ "0.6": 7493,
94
+ "0.7": 6077,
95
+ "0.8": 3854,
96
+ "0.9": 1170
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 2322,
100
+ "0.6": 2234,
101
+ "0.7": 2022,
102
+ "0.8": 1405,
103
+ "0.9": 494
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 439,
107
+ "0.6": 325,
108
+ "0.7": 145,
109
+ "0.8": 50,
110
+ "0.9": 10
111
+ },
112
+ "babel_mu1_signed": 0.6654933094978333,
113
+ "n_dup_abs": {
114
+ "0.5": 8570,
115
+ "0.6": 7525,
116
+ "0.7": 6107,
117
+ "0.8": 3874,
118
+ "0.9": 1172
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 2326,
122
+ "0.6": 2245,
123
+ "0.7": 2036,
124
+ "0.8": 1414,
125
+ "0.9": 494
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 235,
129
+ "0.6": 173,
130
+ "0.7": 147,
131
+ "0.8": 50,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_abs": 0.667901337146759
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": null,
140
+ "l0_raw": null,
141
+ "recon_r2": null,
142
+ "l1_per_token": null,
143
+ "stats_source": "mp-excluded",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 0,
148
+ "dense_rate": 0.0,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 8456,
152
+ "0.6": 7442,
153
+ "0.7": 6006,
154
+ "0.8": 3793,
155
+ "0.9": 1148
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 2327,
159
+ "0.6": 2258,
160
+ "0.7": 1995,
161
+ "0.8": 1400,
162
+ "0.9": 485
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 355,
166
+ "0.6": 144,
167
+ "0.7": 77,
168
+ "0.8": 36,
169
+ "0.9": 10
170
+ },
171
+ "babel_mu1_signed": 0.6603737473487854,
172
+ "n_dup_abs": {
173
+ "0.5": 8512,
174
+ "0.6": 7487,
175
+ "0.7": 6050,
176
+ "0.8": 3818,
177
+ "0.9": 1152
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 2346,
181
+ "0.6": 2272,
182
+ "0.7": 2011,
183
+ "0.8": 1409,
184
+ "0.9": 485
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 338,
188
+ "0.6": 286,
189
+ "0.7": 88,
190
+ "0.8": 36,
191
+ "0.9": 10
192
+ },
193
+ "babel_mu1_abs": 0.6630961298942566
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7817156910896301,
199
+ "dcs_abs": 0.7834139466285706,
200
+ "frac_sign_flipped": 0.039388020833333336,
201
+ "alive_source": "stats",
202
+ "n_alive_a": 12288,
203
+ "n_alive_b": 12288,
204
+ "matched_cos_signed": {
205
+ "0.5": 0.7236672043800354,
206
+ "0.6": 0.7293717265129089,
207
+ "0.7": 0.7402892112731934,
208
+ "0.8": 0.7550824880599976,
209
+ "0.9": 0.7738751173019409
210
+ },
211
+ "unique_frac_signed": {
212
+ "0.5": 0.4935709635416667,
213
+ "0.6": 0.57177734375,
214
+ "0.7": 0.6702473958333334,
215
+ "0.8": 0.80078125,
216
+ "0.9": 0.9436442057291666
217
+ },
218
+ "dedup_dcs_signed": {
219
+ "0.5": 0.3571811193493583,
220
+ "0.6": 0.4170382283919025,
221
+ "0.7": 0.4961769160193702,
222
+ "0.8": 0.6046558986417949,
223
+ "0.9": 0.7302627703999557
224
+ },
225
+ "matched_frac_signed": {
226
+ "0.5": 0.3589680989583333,
227
+ "0.6": 0.4026692708333333,
228
+ "0.7": 0.4476725260416667,
229
+ "0.8": 0.482421875,
230
+ "0.9": 0.443359375
231
+ },
232
+ "matched_cos_abs": {
233
+ "0.5": 0.7263489365577698,
234
+ "0.6": 0.732623279094696,
235
+ "0.7": 0.7427051663398743,
236
+ "0.8": 0.7570679783821106,
237
+ "0.9": 0.7756255269050598
238
+ },
239
+ "unique_frac_abs": {
240
+ "0.5": 0.4913330078125,
241
+ "0.6": 0.5696614583333334,
242
+ "0.7": 0.6688232421875,
243
+ "0.8": 0.7998046875,
244
+ "0.9": 0.9434814453125
245
+ },
246
+ "dedup_dcs_abs": {
247
+ "0.5": 0.35687920772033976,
248
+ "0.6": 0.41734724557803327,
249
+ "0.7": 0.4967384773408412,
250
+ "0.8": 0.6055065178661607,
251
+ "0.9": 0.7317882931456552
252
+ },
253
+ "matched_frac_abs": {
254
+ "0.5": 0.3590494791666667,
255
+ "0.6": 0.4029947916666667,
256
+ "0.7": 0.447998046875,
257
+ "0.8": 0.4825032552083333,
258
+ "0.9": 0.443359375
259
+ }
260
+ },
261
+ "seed0\u2194seed2": {
262
+ "dcs_signed": 0.7770745754241943,
263
+ "dcs_abs": 0.7787559628486633,
264
+ "frac_sign_flipped": 0.043701171875,
265
+ "alive_source": "stats",
266
+ "n_alive_a": 12288,
267
+ "n_alive_b": 12288,
268
+ "matched_cos_signed": {
269
+ "0.5": 0.7226144671440125,
270
+ "0.6": 0.7275781631469727,
271
+ "0.7": 0.7358648180961609,
272
+ "0.8": 0.7517416477203369,
273
+ "0.9": 0.7691615223884583
274
+ },
275
+ "unique_frac_signed": {
276
+ "0.5": 0.4968668619791667,
277
+ "0.6": 0.5748291015625,
278
+ "0.7": 0.6720377604166666,
279
+ "0.8": 0.8030598958333334,
280
+ "0.9": 0.9441731770833334
281
+ },
282
+ "dedup_dcs_signed": {
283
+ "0.5": 0.3590431827105931,
284
+ "0.6": 0.41823310183826834,
285
+ "0.7": 0.49452894432276173,
286
+ "0.8": 0.6036935693118721,
287
+ "0.9": 0.726221678283764
288
+ },
289
+ "matched_frac_signed": {
290
+ "0.5": 0.3600260416666667,
291
+ "0.6": 0.4007161458333333,
292
+ "0.7": 0.4436848958333333,
293
+ "0.8": 0.4747721354166667,
294
+ "0.9": 0.4336751302083333
295
+ },
296
+ "matched_cos_abs": {
297
+ "0.5": 0.7251178026199341,
298
+ "0.6": 0.7300339341163635,
299
+ "0.7": 0.737865686416626,
300
+ "0.8": 0.7535355687141418,
301
+ "0.9": 0.7708365321159363
302
+ },
303
+ "unique_frac_abs": {
304
+ "0.5": 0.4945068359375,
305
+ "0.6": 0.5723063151041666,
306
+ "0.7": 0.6701253255208334,
307
+ "0.8": 0.8018798828125,
308
+ "0.9": 0.9439290364583334
309
+ },
310
+ "dedup_dcs_abs": {
311
+ "0.5": 0.35857571025553625,
312
+ "0.6": 0.417803030735134,
313
+ "0.7": 0.4944624833005946,
314
+ "0.8": 0.6042450135355466,
315
+ "0.9": 0.7276149850270789
316
+ },
317
+ "matched_frac_abs": {
318
+ "0.5": 0.3597005208333333,
319
+ "0.6": 0.4002278645833333,
320
+ "0.7": 0.443115234375,
321
+ "0.8": 0.475830078125,
322
+ "0.9": 0.4337565104166667
323
+ }
324
+ },
325
+ "seed1\u2194seed2": {
326
+ "dcs_signed": 0.7794721722602844,
327
+ "dcs_abs": 0.781230628490448,
328
+ "frac_sign_flipped": 0.042073567708333336,
329
+ "alive_source": "stats",
330
+ "n_alive_a": 12288,
331
+ "n_alive_b": 12288,
332
+ "matched_cos_signed": {
333
+ "0.5": 0.722134530544281,
334
+ "0.6": 0.7282166481018066,
335
+ "0.7": 0.7373058795928955,
336
+ "0.8": 0.7540035247802734,
337
+ "0.9": 0.7709962725639343
338
+ },
339
+ "unique_frac_signed": {
340
+ "0.5": 0.4979248046875,
341
+ "0.6": 0.5750732421875,
342
+ "0.7": 0.6717936197916666,
343
+ "0.8": 0.802978515625,
344
+ "0.9": 0.9455159505208334
345
+ },
346
+ "dedup_dcs_signed": {
347
+ "0.5": 0.3595686950793606,
348
+ "0.6": 0.4187779088388197,
349
+ "0.7": 0.49531738574538997,
350
+ "0.8": 0.6054486311040819,
351
+ "0.9": 0.7289892735013079
352
+ },
353
+ "matched_frac_signed": {
354
+ "0.5": 0.3595377604166667,
355
+ "0.6": 0.3990885416666667,
356
+ "0.7": 0.4423828125,
357
+ "0.8": 0.47998046875,
358
+ "0.9": 0.4520670572916667
359
+ },
360
+ "matched_cos_abs": {
361
+ "0.5": 0.7249367237091064,
362
+ "0.6": 0.7299293875694275,
363
+ "0.7": 0.7390979528427124,
364
+ "0.8": 0.7557072639465332,
365
+ "0.9": 0.7726919054985046
366
+ },
367
+ "unique_frac_abs": {
368
+ "0.5": 0.4950358072916667,
369
+ "0.6": 0.5729573567708334,
370
+ "0.7": 0.6700032552083334,
371
+ "0.8": 0.8018798828125,
372
+ "0.9": 0.9452718098958334
373
+ },
374
+ "dedup_dcs_abs": {
375
+ "0.5": 0.3588696362567134,
376
+ "0.6": 0.41821841253113234,
377
+ "0.7": 0.4951980343224326,
378
+ "0.8": 0.605986452254001,
379
+ "0.9": 0.7304038760024317
380
+ },
381
+ "matched_frac_abs": {
382
+ "0.5": 0.359130859375,
383
+ "0.6": 0.3982747395833333,
384
+ "0.7": 0.442138671875,
385
+ "0.8": 0.48046875,
386
+ "0.9": 0.452392578125
387
+ }
388
+ }
389
+ },
390
+ "summary": {
391
+ "n_seeds": 3,
392
+ "n_pairs": 3,
393
+ "report_band": [
394
+ 0.6,
395
+ 0.7,
396
+ 0.8,
397
+ 0.9
398
+ ],
399
+ "n_dead_mean": 0,
400
+ "n_dead_std": 0.0,
401
+ "n_dense_mean": 0,
402
+ "n_dense_std": 0.0,
403
+ "l0_mean": null,
404
+ "l0_std": null,
405
+ "recon_r2_mean": null,
406
+ "recon_r2_std": null,
407
+ "l1_per_token_mean": null,
408
+ "l1_per_token_std": null,
409
+ "babel_mu1_signed_mean": 0.6643020113309225,
410
+ "babel_mu1_signed_std": 0.0034886552645365565,
411
+ "n_dup_signed_per_threshold": {
412
+ "0.5": 8531.666666666666,
413
+ "0.6": 7498.0,
414
+ "0.7": 6044.666666666667,
415
+ "0.8": 3833.6666666666665,
416
+ "0.9": 1182.0
417
+ },
418
+ "n_dup_signed_band": 4639.583333333334,
419
+ "n_dup_signed_band_std": 38.65418261111381,
420
+ "n_dup_clusters_signed_per_threshold": {
421
+ "0.5": 2340.0,
422
+ "0.6": 2262.0,
423
+ "0.7": 2006.3333333333333,
424
+ "0.8": 1404.0,
425
+ "0.9": 499.3333333333333
426
+ },
427
+ "n_dup_clusters_signed_band": 1542.9166666666665,
428
+ "n_dup_clusters_signed_band_std": 11.10273990208423,
429
+ "largest_dup_signed_per_threshold": {
430
+ "0.5": 387.0,
431
+ "0.6": 235.0,
432
+ "0.7": 118.66666666666667,
433
+ "0.8": 38.333333333333336,
434
+ "0.9": 10.0
435
+ },
436
+ "largest_dup_signed_band": 100.5,
437
+ "largest_dup_signed_band_std": 32.90991491936739,
438
+ "dcs_signed_mean": 0.779420812924703,
439
+ "dcs_signed_std": 0.0023209840566022644,
440
+ "dedup_dcs_signed_per_threshold": {
441
+ "0.5": 0.35859766571310403,
442
+ "0.6": 0.41801641302299686,
443
+ "0.7": 0.49534108202917393,
444
+ "0.8": 0.6045993663525829,
445
+ "0.9": 0.7284912407283425
446
+ },
447
+ "dedup_dcs_signed_band": 0.561612025533274,
448
+ "dedup_dcs_signed_band_std": 0.000817928943294489,
449
+ "matched_cos_signed_per_threshold": {
450
+ "0.5": 0.722805400689443,
451
+ "0.6": 0.7283888459205627,
452
+ "0.7": 0.7378199696540833,
453
+ "0.8": 0.7536092201868693,
454
+ "0.9": 0.7713443040847778
455
+ },
456
+ "matched_cos_signed_band": 0.7477905849615732,
457
+ "matched_cos_signed_band_std": 0.001789422150281412,
458
+ "unique_frac_signed_per_threshold": {
459
+ "0.5": 0.49612087673611116,
460
+ "0.6": 0.5738932291666666,
461
+ "0.7": 0.6713595920138888,
462
+ "0.8": 0.8022732204861112,
463
+ "0.9": 0.9444444444444445
464
+ },
465
+ "unique_frac_signed_band": 0.7479926215277778,
466
+ "unique_frac_signed_band_std": 0.0012055337535602764,
467
+ "matched_frac_signed_per_threshold": {
468
+ "0.5": 0.3595106336805556,
469
+ "0.6": 0.40082465277777773,
470
+ "0.7": 0.444580078125,
471
+ "0.8": 0.47905815972222227,
472
+ "0.9": 0.4430338541666667
473
+ },
474
+ "matched_frac_signed_band": 0.4418741861979167,
475
+ "matched_frac_signed_band_std": 0.003188141718741717,
476
+ "babel_mu1_abs_mean": 0.6668546994527181,
477
+ "babel_mu1_abs_std": 0.0033598266343399697,
478
+ "n_dup_abs_per_threshold": {
479
+ "0.5": 8575.0,
480
+ "0.6": 7539.666666666667,
481
+ "0.7": 6079.333333333333,
482
+ "0.8": 3855.3333333333335,
483
+ "0.9": 1184.0
484
+ },
485
+ "n_dup_abs_band": 4664.583333333333,
486
+ "n_dup_abs_band_std": 35.63033585771166,
487
+ "n_dup_clusters_abs_per_threshold": {
488
+ "0.5": 2352.6666666666665,
489
+ "0.6": 2276.0,
490
+ "0.7": 2020.0,
491
+ "0.8": 1412.3333333333333,
492
+ "0.9": 498.6666666666667
493
+ },
494
+ "n_dup_clusters_abs_band": 1551.75,
495
+ "n_dup_clusters_abs_band_std": 10.5,
496
+ "largest_dup_abs_per_threshold": {
497
+ "0.5": 315.0,
498
+ "0.6": 251.33333333333334,
499
+ "0.7": 116.33333333333333,
500
+ "0.8": 38.333333333333336,
501
+ "0.9": 10.0
502
+ },
503
+ "largest_dup_abs_band": 104.0,
504
+ "largest_dup_abs_band_std": 8.54400374531753,
505
+ "dcs_abs_mean": 0.7811335126558939,
506
+ "dcs_abs_std": 0.0023305099951428116,
507
+ "dedup_dcs_abs_per_threshold": {
508
+ "0.5": 0.3581081847441965,
509
+ "0.6": 0.4177895629480999,
510
+ "0.7": 0.49546633165462284,
511
+ "0.8": 0.6052459945519028,
512
+ "0.9": 0.7299357180583886
513
+ },
514
+ "dedup_dcs_abs_band": 0.5621094018032535,
515
+ "dedup_dcs_abs_band_std": 0.0009540964043521437,
516
+ "matched_cos_abs_per_threshold": {
517
+ "0.5": 0.7254678209622701,
518
+ "0.6": 0.7308622002601624,
519
+ "0.7": 0.7398896018664042,
520
+ "0.8": 0.7554369370142618,
521
+ "0.9": 0.7730513215065002
522
+ },
523
+ "matched_cos_abs_band": 0.7498100151618321,
524
+ "matched_cos_abs_band_std": 0.0020075506852886187,
525
+ "unique_frac_abs_per_threshold": {
526
+ "0.5": 0.4936252170138889,
527
+ "0.6": 0.5716417100694445,
528
+ "0.7": 0.669650607638889,
529
+ "0.8": 0.8011881510416666,
530
+ "0.9": 0.9442274305555557
531
+ },
532
+ "unique_frac_abs_band": 0.746676974826389,
533
+ "unique_frac_abs_band_std": 0.0010942126864736965,
534
+ "matched_frac_abs_per_threshold": {
535
+ "0.5": 0.3592936197916667,
536
+ "0.6": 0.4004991319444444,
537
+ "0.7": 0.4444173177083333,
538
+ "0.8": 0.4796006944444444,
539
+ "0.9": 0.44316948784722227
540
+ },
541
+ "matched_frac_abs_band": 0.4419216579861111,
542
+ "matched_frac_abs_band_std": 0.003226171902808814,
543
+ "alive_source": "stats"
544
+ }
545
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed0/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1298,
3
+ "r2_score": 0.9648,
4
+ "l0_per_token_0": 0.0029,
5
+ "l0_per_token_0.001": 0.0029,
6
+ "l0_per_token_0.01": 0.0029,
7
+ "l0_per_token_0.05": 0.0029,
8
+ "l0_per_token_0.1": 0.0029,
9
+ "l0_per_token_0.5": 0.0029,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 96.7825
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed1/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1258,
3
+ "r2_score": 0.9659,
4
+ "l0_per_token_0": 0.0031,
5
+ "l0_per_token_0.001": 0.0031,
6
+ "l0_per_token_0.01": 0.0031,
7
+ "l0_per_token_0.05": 0.0031,
8
+ "l0_per_token_0.1": 0.0031,
9
+ "l0_per_token_0.5": 0.0031,
10
+ "l0_per_token_1.0": 0.0027,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 100.9962
15
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed2/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "k_iters": 10,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 10,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 1228,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1287,
3
+ "r2_score": 0.9651,
4
+ "l0_per_token_0": 0.003,
5
+ "l0_per_token_0.001": 0.003,
6
+ "l0_per_token_0.01": 0.003,
7
+ "l0_per_token_0.05": 0.003,
8
+ "l0_per_token_0.1": 0.003,
9
+ "l0_per_token_0.5": 0.003,
10
+ "l0_per_token_1.0": 0.0026,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 98.9773
15
+ }
vision/dinov2-b_layer10/mp/x4/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 94.26257976653696,
20
+ "recon_r2": 0.975821934346816,
21
+ "l1_per_token": 160.12352198927357
22
+ },
23
+ "seed1": {
24
+ "l0": 96.0824747081712,
25
+ "recon_r2": 0.9758979026055418,
26
+ "l1_per_token": 160.55607186960745
27
+ },
28
+ "seed2": {
29
+ "l0": 96.16374837872893,
30
+ "recon_r2": 0.9758934109302257,
31
+ "l1_per_token": 160.4021775571133
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.686996152798104,
38
+ "mcc_abs": 0.6870703592996316,
39
+ "act_iou": 0.25547694536790305,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9846564578689908,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8233672574692353,
44
+ "0.7": 0.8235882716727134,
45
+ "0.8": 0.8303783596636519,
46
+ "0.9": 0.8476126278762411
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8221728981196199,
50
+ "0.7": 0.822716947685233,
51
+ "0.8": 0.831733380016577,
52
+ "0.9": 0.8478334868152545
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.6846840780527694,
58
+ "mcc_abs": 0.6847622419526243,
59
+ "act_iou": 0.25246337038642785,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9844104815350083,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8227527574304134,
64
+ "0.7": 0.8178582045636273,
65
+ "0.8": 0.8274592190546871,
66
+ "0.9": 0.844338840962079
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8202325549355041,
70
+ "0.7": 0.8191295136242774,
71
+ "0.8": 0.8274661268153322,
72
+ "0.9": 0.8450663089562113
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.6946268447282802,
78
+ "mcc_abs": 0.6946844550087684,
79
+ "act_iou": 0.2577866682674802,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9846777056136713,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.822279493121979,
84
+ "0.7": 0.818058897473975,
85
+ "0.8": 0.8203782644254983,
86
+ "0.9": 0.8269563531697791
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.820086935363402,
90
+ "0.7": 0.8187139526860329,
91
+ "0.8": 0.819773872370219,
92
+ "0.9": 0.8279184476094364
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 95.50293428447902,
106
+ "l0_std": 1.0749469056730019,
107
+ "recon_r2_mean": 0.975871082627528,
108
+ "recon_r2_std": 4.2622868386407093e-05,
109
+ "l1_per_token_mean": 160.3605904719981,
110
+ "l1_per_token_std": 0.2192531960553768,
111
+ "input_r2_mean": 0.9845815483392236,
112
+ "input_r2_std": 0.00014852863458536,
113
+ "mcc_signed_mean": 0.6887690251930513,
114
+ "mcc_signed_std": 0.005203072139369793,
115
+ "mcc_abs_mean": 0.6888390187536748,
116
+ "mcc_abs_std": 0.005192176343419167,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8227998360072094,
119
+ "0.7": 0.8198351245701052,
120
+ "0.8": 0.8260719477146125,
121
+ "0.9": 0.8396359406693664
122
+ },
123
+ "matched_mcc_signed_band": 0.8270857122403233,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8208307961395086,
126
+ "0.7": 0.8201868046651811,
127
+ "0.8": 0.8263244597340428,
128
+ "0.9": 0.8402727477936341
129
+ },
130
+ "matched_mcc_abs_band": 0.8269037020830917,
131
+ "act_iou_mean": 0.25524232800727037,
132
+ "act_iou_std": 0.002669393013783618,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x4/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,595 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 3,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 633,
34
+ "0.6": 505,
35
+ "0.7": 383,
36
+ "0.8": 264,
37
+ "0.9": 129
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 178,
41
+ "0.6": 140,
42
+ "0.7": 110,
43
+ "0.8": 71,
44
+ "0.9": 33
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 455,
48
+ "0.6": 365,
49
+ "0.7": 273,
50
+ "0.8": 193,
51
+ "0.9": 96
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 102,
55
+ "0.6": 56,
56
+ "0.7": 46,
57
+ "0.8": 27,
58
+ "0.9": 14
59
+ },
60
+ "babel_mu1_signed": 0.3663797080516815,
61
+ "n_dup_abs": {
62
+ "0.5": 644,
63
+ "0.6": 511,
64
+ "0.7": 389,
65
+ "0.8": 267,
66
+ "0.9": 130
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 181,
70
+ "0.6": 142,
71
+ "0.7": 112,
72
+ "0.8": 72,
73
+ "0.9": 34
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 463,
77
+ "0.6": 369,
78
+ "0.7": 277,
79
+ "0.8": 195,
80
+ "0.9": 96
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 106,
84
+ "0.6": 55,
85
+ "0.7": 51,
86
+ "0.8": 27,
87
+ "0.9": 11
88
+ },
89
+ "babel_mu1_abs": 0.375279039144516,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": null,
96
+ "l0_raw": null,
97
+ "recon_r2": null,
98
+ "l1_per_token": null,
99
+ "stats_source": "mp-excluded",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 0,
104
+ "dense_rate": 0.0,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 658,
108
+ "0.6": 541,
109
+ "0.7": 413,
110
+ "0.8": 303,
111
+ "0.9": 198
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 174,
115
+ "0.6": 140,
116
+ "0.7": 100,
117
+ "0.8": 65,
118
+ "0.9": 60
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 484,
122
+ "0.6": 401,
123
+ "0.7": 313,
124
+ "0.8": 238,
125
+ "0.9": 138
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 89,
129
+ "0.6": 48,
130
+ "0.7": 28,
131
+ "0.8": 22,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_signed": 0.3705531060695648,
135
+ "n_dup_abs": {
136
+ "0.5": 671,
137
+ "0.6": 548,
138
+ "0.7": 414,
139
+ "0.8": 308,
140
+ "0.9": 200
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 179,
144
+ "0.6": 140,
145
+ "0.7": 97,
146
+ "0.8": 67,
147
+ "0.9": 61
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 492,
151
+ "0.6": 408,
152
+ "0.7": 317,
153
+ "0.8": 241,
154
+ "0.9": 139
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 73,
158
+ "0.6": 70,
159
+ "0.7": 28,
160
+ "0.8": 22,
161
+ "0.9": 10
162
+ },
163
+ "babel_mu1_abs": 0.3797185719013214,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": null,
170
+ "l0_raw": null,
171
+ "recon_r2": null,
172
+ "l1_per_token": null,
173
+ "stats_source": "mp-excluded",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 0,
178
+ "dense_rate": 0.0,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 639,
182
+ "0.6": 540,
183
+ "0.7": 423,
184
+ "0.8": 318,
185
+ "0.9": 208
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 156,
189
+ "0.6": 128,
190
+ "0.7": 99,
191
+ "0.8": 74,
192
+ "0.9": 63
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 483,
196
+ "0.6": 412,
197
+ "0.7": 324,
198
+ "0.8": 244,
199
+ "0.9": 145
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 111,
203
+ "0.6": 57,
204
+ "0.7": 56,
205
+ "0.8": 45,
206
+ "0.9": 14
207
+ },
208
+ "babel_mu1_signed": 0.3651764392852783,
209
+ "n_dup_abs": {
210
+ "0.5": 642,
211
+ "0.6": 548,
212
+ "0.7": 429,
213
+ "0.8": 324,
214
+ "0.9": 212
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 158,
218
+ "0.6": 133,
219
+ "0.7": 103,
220
+ "0.8": 78,
221
+ "0.9": 65
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 484,
225
+ "0.6": 415,
226
+ "0.7": 326,
227
+ "0.8": 246,
228
+ "0.9": 147
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 62,
232
+ "0.6": 57,
233
+ "0.7": 56,
234
+ "0.8": 45,
235
+ "0.9": 14
236
+ },
237
+ "babel_mu1_abs": 0.37334656715393066,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.863697350025177,
244
+ "dcs_abs": 0.8660604953765869,
245
+ "frac_sign_flipped": 0.033528645833333336,
246
+ "alive_source": "stats",
247
+ "frac_denominator": "d_sae",
248
+ "n_alive_a": 3072,
249
+ "n_alive_b": 3072,
250
+ "matched_cos_signed": {
251
+ "0.5": 0.8904803991317749,
252
+ "0.6": 0.8863230347633362,
253
+ "0.7": 0.8811215758323669,
254
+ "0.8": 0.8779223561286926,
255
+ "0.9": 0.8697978258132935
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.84716796875,
259
+ "0.6": 0.8753255208333334,
260
+ "0.7": 0.9046223958333334,
261
+ "0.8": 0.9298502604166666,
262
+ "0.9": 0.9619140625
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.754386470944155,
266
+ "0.6": 0.7758211720307979,
267
+ "0.7": 0.7970823109499179,
268
+ "0.8": 0.8163363314718783,
269
+ "0.9": 0.8366707601817325
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.7802734375,
273
+ "0.6": 0.7838541666666666,
274
+ "0.7": 0.7750651041666666,
275
+ "0.8": 0.7529296875,
276
+ "0.9": 0.6888020833333334
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.892687201499939,
280
+ "0.6": 0.8885923027992249,
281
+ "0.7": 0.8832919001579285,
282
+ "0.8": 0.8800186514854431,
283
+ "0.9": 0.8724757432937622
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.8445638020833334,
287
+ "0.6": 0.87353515625,
288
+ "0.7": 0.9033203125,
289
+ "0.8": 0.9290364583333334,
290
+ "0.9": 0.9617513020833334
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.7539312969699192,
294
+ "0.6": 0.7762166160682682,
295
+ "0.7": 0.7978955152793787,
296
+ "0.8": 0.8175694112433121,
297
+ "0.9": 0.8391046821488999
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.7802734375,
301
+ "0.6": 0.7848307291666666,
302
+ "0.7": 0.7760416666666666,
303
+ "0.8": 0.751953125,
304
+ "0.9": 0.6884765625
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.8620738983154297,
309
+ "dcs_abs": 0.8638548851013184,
310
+ "frac_sign_flipped": 0.028645833333333332,
311
+ "alive_source": "stats",
312
+ "frac_denominator": "d_sae",
313
+ "n_alive_a": 3072,
314
+ "n_alive_b": 3072,
315
+ "matched_cos_signed": {
316
+ "0.5": 0.8867824077606201,
317
+ "0.6": 0.8854392170906067,
318
+ "0.7": 0.8801518082618713,
319
+ "0.8": 0.8756371140480042,
320
+ "0.9": 0.8707349896430969
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.8473307291666666,
324
+ "0.6": 0.87353515625,
325
+ "0.7": 0.90283203125,
326
+ "0.8": 0.9288736979166666,
327
+ "0.9": 0.9607747395833334
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.7513979841799785,
331
+ "0.6": 0.7734622848511208,
332
+ "0.7": 0.7946292448614258,
333
+ "0.8": 0.8133562841588475,
334
+ "0.9": 0.8365801829204429
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.7734375,
338
+ "0.6": 0.7809244791666666,
339
+ "0.7": 0.7770182291666666,
340
+ "0.8": 0.7513020833333334,
341
+ "0.9": 0.6901041666666666
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.8874406218528748,
345
+ "0.6": 0.8861882090568542,
346
+ "0.7": 0.8809804916381836,
347
+ "0.8": 0.8767107129096985,
348
+ "0.9": 0.8725345730781555
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.8458658854166666,
352
+ "0.6": 0.8723958333333334,
353
+ "0.7": 0.90185546875,
354
+ "0.8": 0.92822265625,
355
+ "0.9": 0.96044921875
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.7506557473582991,
359
+ "0.6": 0.7731069011303286,
360
+ "0.7": 0.7945170742459595,
361
+ "0.8": 0.8137827466998715,
362
+ "0.9": 0.8380251490452792
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.7737630208333334,
366
+ "0.6": 0.7822265625,
367
+ "0.7": 0.7760416666666666,
368
+ "0.8": 0.7506510416666666,
369
+ "0.9": 0.6888020833333334
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8889439702033997,
374
+ "dcs_abs": 0.8901732563972473,
375
+ "frac_sign_flipped": 0.0185546875,
376
+ "alive_source": "stats",
377
+ "frac_denominator": "d_sae",
378
+ "n_alive_a": 3072,
379
+ "n_alive_b": 3072,
380
+ "matched_cos_signed": {
381
+ "0.5": 0.8958790898323059,
382
+ "0.6": 0.893430769443512,
383
+ "0.7": 0.889135479927063,
384
+ "0.8": 0.8864941000938416,
385
+ "0.9": 0.8865686058998108
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.8426106770833334,
389
+ "0.6": 0.86767578125,
390
+ "0.7": 0.8963216145833334,
391
+ "0.8": 0.9215494791666666,
392
+ "0.9": 0.9539388020833334
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.7548772864683997,
396
+ "0.6": 0.7752082408696879,
397
+ "0.7": 0.7969513489515521,
398
+ "0.8": 0.8169481762258025,
399
+ "0.9": 0.8457321938767564
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.7760416666666666,
403
+ "0.6": 0.7858072916666666,
404
+ "0.7": 0.78515625,
405
+ "0.8": 0.7705078125,
406
+ "0.9": 0.7272135416666666
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.8988944888114929,
410
+ "0.6": 0.8942874073982239,
411
+ "0.7": 0.8906096816062927,
412
+ "0.8": 0.8878304362297058,
413
+ "0.9": 0.8880638480186462
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.8411458333333334,
417
+ "0.6": 0.8660481770833334,
418
+ "0.7": 0.8953450520833334,
419
+ "0.8": 0.9207356770833334,
420
+ "0.9": 0.9534505208333334
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.7561013538700839,
424
+ "0.6": 0.7744959789658121,
425
+ "0.7": 0.7974029717637071,
426
+ "0.8": 0.8174571578371495,
427
+ "0.9": 0.8467249384266324
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.77734375,
431
+ "0.6": 0.787109375,
432
+ "0.7": 0.7854817708333334,
433
+ "0.8": 0.7708333333333334,
434
+ "0.9": 0.7259114583333334
435
+ }
436
+ }
437
+ },
438
+ "summary": {
439
+ "n_seeds": 3,
440
+ "n_pairs": 3,
441
+ "report_band": [
442
+ 0.6,
443
+ 0.7,
444
+ 0.8,
445
+ 0.9
446
+ ],
447
+ "n_dead_mean": 0,
448
+ "n_dead_std": 0.0,
449
+ "n_dense_mean": 0,
450
+ "n_dense_std": 0.0,
451
+ "l0_mean": null,
452
+ "l0_std": null,
453
+ "recon_r2_mean": null,
454
+ "recon_r2_std": null,
455
+ "l1_per_token_mean": null,
456
+ "l1_per_token_std": null,
457
+ "babel_mu1_signed_mean": 0.36736975113550824,
458
+ "babel_mu1_signed_std": 0.0028217504156440275,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 643.3333333333334,
461
+ "0.6": 528.6666666666666,
462
+ "0.7": 406.3333333333333,
463
+ "0.8": 295.0,
464
+ "0.9": 178.33333333333334
465
+ },
466
+ "n_dup_signed_band": 352.0833333333333,
467
+ "n_dup_signed_band_std": 27.894145144336893,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 169.33333333333334,
470
+ "0.6": 136.0,
471
+ "0.7": 103.0,
472
+ "0.8": 70.0,
473
+ "0.9": 52.0
474
+ },
475
+ "n_dup_clusters_signed_band": 90.25,
476
+ "n_dup_clusters_signed_band_std": 1.5206906325745548,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 100.66666666666667,
479
+ "0.6": 53.666666666666664,
480
+ "0.7": 43.333333333333336,
481
+ "0.8": 31.333333333333332,
482
+ "0.9": 12.666666666666666
483
+ },
484
+ "largest_dup_signed_band": 35.25,
485
+ "largest_dup_signed_band_std": 8.011710179481033,
486
+ "dcs_signed_mean": 0.8715717395146688,
487
+ "dcs_signed_std": 0.015066675087434934,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.7535539138641777,
490
+ "0.6": 0.7748305659172022,
491
+ "0.7": 0.7962209682542986,
492
+ "0.8": 0.8155469306188428,
493
+ "0.9": 0.839661045659644
494
+ },
495
+ "dedup_dcs_signed_band": 0.8065648776124968,
496
+ "dedup_dcs_signed_band_std": 0.002102852872293367,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.8910472989082336,
499
+ "0.6": 0.8883976737658182,
500
+ "0.7": 0.8834696213404337,
501
+ "0.8": 0.8800178567568461,
502
+ "0.9": 0.8757004737854004
503
+ },
504
+ "matched_cos_signed_band": 0.8818964064121246,
505
+ "matched_cos_signed_band_std": 0.006084734579468042,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.845703125,
508
+ "0.6": 0.8721788194444445,
509
+ "0.7": 0.9012586805555557,
510
+ "0.8": 0.9267578125,
511
+ "0.9": 0.9588758680555557
512
+ },
513
+ "unique_frac_signed_band": 0.914767795138889,
514
+ "unique_frac_signed_band_std": 0.004299758816805034,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.7765842013888888,
517
+ "0.6": 0.7835286458333334,
518
+ "0.7": 0.779079861111111,
519
+ "0.8": 0.7582465277777778,
520
+ "0.9": 0.7020399305555555
521
+ },
522
+ "matched_frac_signed_band": 0.7557237413194444,
523
+ "matched_frac_signed_band_std": 0.009915146747727952,
524
+ "babel_mu1_abs_mean": 0.37611472606658936,
525
+ "babel_mu1_abs_std": 0.003267168590536138,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 652.3333333333334,
528
+ "0.6": 535.6666666666666,
529
+ "0.7": 410.6666666666667,
530
+ "0.8": 299.6666666666667,
531
+ "0.9": 180.66666666666666
532
+ },
533
+ "n_dup_abs_band": 356.6666666666667,
534
+ "n_dup_abs_band_std": 28.583576286625387,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 172.66666666666666,
537
+ "0.6": 138.33333333333334,
538
+ "0.7": 104.0,
539
+ "0.8": 72.33333333333333,
540
+ "0.9": 53.333333333333336
541
+ },
542
+ "n_dup_clusters_abs_band": 92.0,
543
+ "n_dup_clusters_abs_band_std": 2.462214450449026,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 80.33333333333333,
546
+ "0.6": 60.666666666666664,
547
+ "0.7": 45.0,
548
+ "0.8": 31.333333333333332,
549
+ "0.9": 11.666666666666666
550
+ },
551
+ "largest_dup_abs_band": 37.166666666666664,
552
+ "largest_dup_abs_band_std": 5.346338310781814,
553
+ "dcs_abs_mean": 0.8733628789583842,
554
+ "dcs_abs_std": 0.014599923677861167,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.7535627993994342,
557
+ "0.6": 0.7746064987214697,
558
+ "0.7": 0.7966051870963485,
559
+ "0.8": 0.8162697719267777,
560
+ "0.9": 0.8412849232069372
561
+ },
562
+ "dedup_dcs_abs_band": 0.8071915952378833,
563
+ "dedup_dcs_abs_band_std": 0.002126596326879785,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.8930074373881022,
566
+ "0.6": 0.889689306418101,
567
+ "0.7": 0.8849606911341349,
568
+ "0.8": 0.8815199335416158,
569
+ "0.9": 0.877691388130188
570
+ },
571
+ "matched_cos_abs_band": 0.8834653298060099,
572
+ "matched_cos_abs_band_std": 0.005914915546760715,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.8438585069444445,
575
+ "0.6": 0.8706597222222223,
576
+ "0.7": 0.9001736111111112,
577
+ "0.8": 0.925998263888889,
578
+ "0.9": 0.9585503472222223
579
+ },
580
+ "unique_frac_abs_band": 0.9138454861111112,
581
+ "unique_frac_abs_band_std": 0.00432777723906263,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.7771267361111112,
584
+ "0.6": 0.7847222222222222,
585
+ "0.7": 0.7791883680555555,
586
+ "0.8": 0.7578125,
587
+ "0.9": 0.7010633680555557
588
+ },
589
+ "matched_frac_abs_band": 0.7556966145833334,
590
+ "matched_frac_abs_band_std": 0.01008819208703062,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
vision/dinov2-b_layer10/mp/x4/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 94.26257976653696,
20
+ "recon_r2": 0.975821934346816,
21
+ "l1_per_token": 160.12352198927357
22
+ },
23
+ "seed1": {
24
+ "l0": 96.0824747081712,
25
+ "recon_r2": 0.9758979026055418,
26
+ "l1_per_token": 160.55607186960745
27
+ },
28
+ "seed2": {
29
+ "l0": 96.16374837872893,
30
+ "recon_r2": 0.9758934109302257,
31
+ "l1_per_token": 160.4021775571133
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.686996152798104,
38
+ "mcc_abs": 0.6870703592996316,
39
+ "act_iou": 0.25547694536790305,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9846564578689908,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8233672574692353,
44
+ "0.7": 0.8235882716727134,
45
+ "0.8": 0.8303783596636519,
46
+ "0.9": 0.8476126278762411
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8221728981196199,
50
+ "0.7": 0.822716947685233,
51
+ "0.8": 0.831733380016577,
52
+ "0.9": 0.8478334868152545
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.6846840780527694,
58
+ "mcc_abs": 0.6847622419526243,
59
+ "act_iou": 0.25246337038642785,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9844104815350083,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8227527574304134,
64
+ "0.7": 0.8178582045636273,
65
+ "0.8": 0.8274592190546871,
66
+ "0.9": 0.844338840962079
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8202325549355041,
70
+ "0.7": 0.8191295136242774,
71
+ "0.8": 0.8274661268153322,
72
+ "0.9": 0.8450663089562113
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.6946268447282802,
78
+ "mcc_abs": 0.6946844550087684,
79
+ "act_iou": 0.2577866682674802,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9846777056136713,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.822279493121979,
84
+ "0.7": 0.818058897473975,
85
+ "0.8": 0.8203782644254983,
86
+ "0.9": 0.8269563531697791
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.820086935363402,
90
+ "0.7": 0.8187139526860329,
91
+ "0.8": 0.819773872370219,
92
+ "0.9": 0.8279184476094364
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 95.50293428447902,
106
+ "l0_std": 1.0749469056730019,
107
+ "recon_r2_mean": 0.975871082627528,
108
+ "recon_r2_std": 4.2622868386407093e-05,
109
+ "l1_per_token_mean": 160.3605904719981,
110
+ "l1_per_token_std": 0.2192531960553768,
111
+ "input_r2_mean": 0.9845815483392236,
112
+ "input_r2_std": 0.00014852863458536,
113
+ "mcc_signed_mean": 0.6887690251930513,
114
+ "mcc_signed_std": 0.005203072139369793,
115
+ "mcc_abs_mean": 0.6888390187536748,
116
+ "mcc_abs_std": 0.005192176343419167,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8227998360072094,
119
+ "0.7": 0.8198351245701052,
120
+ "0.8": 0.8260719477146125,
121
+ "0.9": 0.8396359406693664
122
+ },
123
+ "matched_mcc_signed_band": 0.8270857122403233,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8208307961395086,
126
+ "0.7": 0.8201868046651811,
127
+ "0.8": 0.8263244597340428,
128
+ "0.9": 0.8402727477936341
129
+ },
130
+ "matched_mcc_abs_band": 0.8269037020830917,
131
+ "act_iou_mean": 0.25524232800727037,
132
+ "act_iou_std": 0.002669393013783618,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x4/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,595 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 3,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 633,
34
+ "0.6": 505,
35
+ "0.7": 383,
36
+ "0.8": 264,
37
+ "0.9": 129
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 178,
41
+ "0.6": 140,
42
+ "0.7": 110,
43
+ "0.8": 71,
44
+ "0.9": 33
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 455,
48
+ "0.6": 365,
49
+ "0.7": 273,
50
+ "0.8": 193,
51
+ "0.9": 96
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 102,
55
+ "0.6": 56,
56
+ "0.7": 46,
57
+ "0.8": 27,
58
+ "0.9": 14
59
+ },
60
+ "babel_mu1_signed": 0.3663797080516815,
61
+ "n_dup_abs": {
62
+ "0.5": 644,
63
+ "0.6": 511,
64
+ "0.7": 389,
65
+ "0.8": 267,
66
+ "0.9": 130
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 181,
70
+ "0.6": 142,
71
+ "0.7": 112,
72
+ "0.8": 72,
73
+ "0.9": 34
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 463,
77
+ "0.6": 369,
78
+ "0.7": 277,
79
+ "0.8": 195,
80
+ "0.9": 96
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 106,
84
+ "0.6": 55,
85
+ "0.7": 51,
86
+ "0.8": 27,
87
+ "0.9": 11
88
+ },
89
+ "babel_mu1_abs": 0.375279039144516,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": null,
96
+ "l0_raw": null,
97
+ "recon_r2": null,
98
+ "l1_per_token": null,
99
+ "stats_source": "mp-excluded",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 0,
104
+ "dense_rate": 0.0,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 658,
108
+ "0.6": 541,
109
+ "0.7": 413,
110
+ "0.8": 303,
111
+ "0.9": 198
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 174,
115
+ "0.6": 140,
116
+ "0.7": 100,
117
+ "0.8": 65,
118
+ "0.9": 60
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 484,
122
+ "0.6": 401,
123
+ "0.7": 313,
124
+ "0.8": 238,
125
+ "0.9": 138
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 89,
129
+ "0.6": 48,
130
+ "0.7": 28,
131
+ "0.8": 22,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_signed": 0.3705531060695648,
135
+ "n_dup_abs": {
136
+ "0.5": 671,
137
+ "0.6": 548,
138
+ "0.7": 414,
139
+ "0.8": 308,
140
+ "0.9": 200
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 179,
144
+ "0.6": 140,
145
+ "0.7": 97,
146
+ "0.8": 67,
147
+ "0.9": 61
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 492,
151
+ "0.6": 408,
152
+ "0.7": 317,
153
+ "0.8": 241,
154
+ "0.9": 139
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 73,
158
+ "0.6": 70,
159
+ "0.7": 28,
160
+ "0.8": 22,
161
+ "0.9": 10
162
+ },
163
+ "babel_mu1_abs": 0.3797185719013214,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": null,
170
+ "l0_raw": null,
171
+ "recon_r2": null,
172
+ "l1_per_token": null,
173
+ "stats_source": "mp-excluded",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 0,
178
+ "dense_rate": 0.0,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 639,
182
+ "0.6": 540,
183
+ "0.7": 423,
184
+ "0.8": 318,
185
+ "0.9": 208
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 156,
189
+ "0.6": 128,
190
+ "0.7": 99,
191
+ "0.8": 74,
192
+ "0.9": 63
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 483,
196
+ "0.6": 412,
197
+ "0.7": 324,
198
+ "0.8": 244,
199
+ "0.9": 145
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 111,
203
+ "0.6": 57,
204
+ "0.7": 56,
205
+ "0.8": 45,
206
+ "0.9": 14
207
+ },
208
+ "babel_mu1_signed": 0.3651764392852783,
209
+ "n_dup_abs": {
210
+ "0.5": 642,
211
+ "0.6": 548,
212
+ "0.7": 429,
213
+ "0.8": 324,
214
+ "0.9": 212
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 158,
218
+ "0.6": 133,
219
+ "0.7": 103,
220
+ "0.8": 78,
221
+ "0.9": 65
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 484,
225
+ "0.6": 415,
226
+ "0.7": 326,
227
+ "0.8": 246,
228
+ "0.9": 147
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 62,
232
+ "0.6": 57,
233
+ "0.7": 56,
234
+ "0.8": 45,
235
+ "0.9": 14
236
+ },
237
+ "babel_mu1_abs": 0.37334656715393066,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.863697350025177,
244
+ "dcs_abs": 0.8660604953765869,
245
+ "frac_sign_flipped": 0.033528645833333336,
246
+ "alive_source": "stats",
247
+ "frac_denominator": "d_sae",
248
+ "n_alive_a": 3072,
249
+ "n_alive_b": 3072,
250
+ "matched_cos_signed": {
251
+ "0.5": 0.8904803991317749,
252
+ "0.6": 0.8863230347633362,
253
+ "0.7": 0.8811215758323669,
254
+ "0.8": 0.8779223561286926,
255
+ "0.9": 0.8697978258132935
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.84716796875,
259
+ "0.6": 0.8753255208333334,
260
+ "0.7": 0.9046223958333334,
261
+ "0.8": 0.9298502604166666,
262
+ "0.9": 0.9619140625
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.754386470944155,
266
+ "0.6": 0.7758211720307979,
267
+ "0.7": 0.7970823109499179,
268
+ "0.8": 0.8163363314718783,
269
+ "0.9": 0.8366707601817325
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.7802734375,
273
+ "0.6": 0.7838541666666666,
274
+ "0.7": 0.7750651041666666,
275
+ "0.8": 0.7529296875,
276
+ "0.9": 0.6888020833333334
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.892687201499939,
280
+ "0.6": 0.8885923027992249,
281
+ "0.7": 0.8832919001579285,
282
+ "0.8": 0.8800186514854431,
283
+ "0.9": 0.8724757432937622
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.8445638020833334,
287
+ "0.6": 0.87353515625,
288
+ "0.7": 0.9033203125,
289
+ "0.8": 0.9290364583333334,
290
+ "0.9": 0.9617513020833334
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.7539312969699192,
294
+ "0.6": 0.7762166160682682,
295
+ "0.7": 0.7978955152793787,
296
+ "0.8": 0.8175694112433121,
297
+ "0.9": 0.8391046821488999
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.7802734375,
301
+ "0.6": 0.7848307291666666,
302
+ "0.7": 0.7760416666666666,
303
+ "0.8": 0.751953125,
304
+ "0.9": 0.6884765625
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.8620738983154297,
309
+ "dcs_abs": 0.8638548851013184,
310
+ "frac_sign_flipped": 0.028645833333333332,
311
+ "alive_source": "stats",
312
+ "frac_denominator": "d_sae",
313
+ "n_alive_a": 3072,
314
+ "n_alive_b": 3072,
315
+ "matched_cos_signed": {
316
+ "0.5": 0.8867824077606201,
317
+ "0.6": 0.8854392170906067,
318
+ "0.7": 0.8801518082618713,
319
+ "0.8": 0.8756371140480042,
320
+ "0.9": 0.8707349896430969
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.8473307291666666,
324
+ "0.6": 0.87353515625,
325
+ "0.7": 0.90283203125,
326
+ "0.8": 0.9288736979166666,
327
+ "0.9": 0.9607747395833334
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.7513979841799785,
331
+ "0.6": 0.7734622848511208,
332
+ "0.7": 0.7946292448614258,
333
+ "0.8": 0.8133562841588475,
334
+ "0.9": 0.8365801829204429
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.7734375,
338
+ "0.6": 0.7809244791666666,
339
+ "0.7": 0.7770182291666666,
340
+ "0.8": 0.7513020833333334,
341
+ "0.9": 0.6901041666666666
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.8874406218528748,
345
+ "0.6": 0.8861882090568542,
346
+ "0.7": 0.8809804916381836,
347
+ "0.8": 0.8767107129096985,
348
+ "0.9": 0.8725345730781555
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.8458658854166666,
352
+ "0.6": 0.8723958333333334,
353
+ "0.7": 0.90185546875,
354
+ "0.8": 0.92822265625,
355
+ "0.9": 0.96044921875
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.7506557473582991,
359
+ "0.6": 0.7731069011303286,
360
+ "0.7": 0.7945170742459595,
361
+ "0.8": 0.8137827466998715,
362
+ "0.9": 0.8380251490452792
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.7737630208333334,
366
+ "0.6": 0.7822265625,
367
+ "0.7": 0.7760416666666666,
368
+ "0.8": 0.7506510416666666,
369
+ "0.9": 0.6888020833333334
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8889439702033997,
374
+ "dcs_abs": 0.8901732563972473,
375
+ "frac_sign_flipped": 0.0185546875,
376
+ "alive_source": "stats",
377
+ "frac_denominator": "d_sae",
378
+ "n_alive_a": 3072,
379
+ "n_alive_b": 3072,
380
+ "matched_cos_signed": {
381
+ "0.5": 0.8958790898323059,
382
+ "0.6": 0.893430769443512,
383
+ "0.7": 0.889135479927063,
384
+ "0.8": 0.8864941000938416,
385
+ "0.9": 0.8865686058998108
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.8426106770833334,
389
+ "0.6": 0.86767578125,
390
+ "0.7": 0.8963216145833334,
391
+ "0.8": 0.9215494791666666,
392
+ "0.9": 0.9539388020833334
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.7548772864683997,
396
+ "0.6": 0.7752082408696879,
397
+ "0.7": 0.7969513489515521,
398
+ "0.8": 0.8169481762258025,
399
+ "0.9": 0.8457321938767564
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.7760416666666666,
403
+ "0.6": 0.7858072916666666,
404
+ "0.7": 0.78515625,
405
+ "0.8": 0.7705078125,
406
+ "0.9": 0.7272135416666666
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.8988944888114929,
410
+ "0.6": 0.8942874073982239,
411
+ "0.7": 0.8906096816062927,
412
+ "0.8": 0.8878304362297058,
413
+ "0.9": 0.8880638480186462
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.8411458333333334,
417
+ "0.6": 0.8660481770833334,
418
+ "0.7": 0.8953450520833334,
419
+ "0.8": 0.9207356770833334,
420
+ "0.9": 0.9534505208333334
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.7561013538700839,
424
+ "0.6": 0.7744959789658121,
425
+ "0.7": 0.7974029717637071,
426
+ "0.8": 0.8174571578371495,
427
+ "0.9": 0.8467249384266324
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.77734375,
431
+ "0.6": 0.787109375,
432
+ "0.7": 0.7854817708333334,
433
+ "0.8": 0.7708333333333334,
434
+ "0.9": 0.7259114583333334
435
+ }
436
+ }
437
+ },
438
+ "summary": {
439
+ "n_seeds": 3,
440
+ "n_pairs": 3,
441
+ "report_band": [
442
+ 0.6,
443
+ 0.7,
444
+ 0.8,
445
+ 0.9
446
+ ],
447
+ "n_dead_mean": 0,
448
+ "n_dead_std": 0.0,
449
+ "n_dense_mean": 0,
450
+ "n_dense_std": 0.0,
451
+ "l0_mean": null,
452
+ "l0_std": null,
453
+ "recon_r2_mean": null,
454
+ "recon_r2_std": null,
455
+ "l1_per_token_mean": null,
456
+ "l1_per_token_std": null,
457
+ "babel_mu1_signed_mean": 0.36736975113550824,
458
+ "babel_mu1_signed_std": 0.0028217504156440275,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 643.3333333333334,
461
+ "0.6": 528.6666666666666,
462
+ "0.7": 406.3333333333333,
463
+ "0.8": 295.0,
464
+ "0.9": 178.33333333333334
465
+ },
466
+ "n_dup_signed_band": 352.0833333333333,
467
+ "n_dup_signed_band_std": 27.894145144336893,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 169.33333333333334,
470
+ "0.6": 136.0,
471
+ "0.7": 103.0,
472
+ "0.8": 70.0,
473
+ "0.9": 52.0
474
+ },
475
+ "n_dup_clusters_signed_band": 90.25,
476
+ "n_dup_clusters_signed_band_std": 1.5206906325745548,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 100.66666666666667,
479
+ "0.6": 53.666666666666664,
480
+ "0.7": 43.333333333333336,
481
+ "0.8": 31.333333333333332,
482
+ "0.9": 12.666666666666666
483
+ },
484
+ "largest_dup_signed_band": 35.25,
485
+ "largest_dup_signed_band_std": 8.011710179481033,
486
+ "dcs_signed_mean": 0.8715717395146688,
487
+ "dcs_signed_std": 0.015066675087434934,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.7535539138641777,
490
+ "0.6": 0.7748305659172022,
491
+ "0.7": 0.7962209682542986,
492
+ "0.8": 0.8155469306188428,
493
+ "0.9": 0.839661045659644
494
+ },
495
+ "dedup_dcs_signed_band": 0.8065648776124968,
496
+ "dedup_dcs_signed_band_std": 0.002102852872293367,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.8910472989082336,
499
+ "0.6": 0.8883976737658182,
500
+ "0.7": 0.8834696213404337,
501
+ "0.8": 0.8800178567568461,
502
+ "0.9": 0.8757004737854004
503
+ },
504
+ "matched_cos_signed_band": 0.8818964064121246,
505
+ "matched_cos_signed_band_std": 0.006084734579468042,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.845703125,
508
+ "0.6": 0.8721788194444445,
509
+ "0.7": 0.9012586805555557,
510
+ "0.8": 0.9267578125,
511
+ "0.9": 0.9588758680555557
512
+ },
513
+ "unique_frac_signed_band": 0.914767795138889,
514
+ "unique_frac_signed_band_std": 0.004299758816805034,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.7765842013888888,
517
+ "0.6": 0.7835286458333334,
518
+ "0.7": 0.779079861111111,
519
+ "0.8": 0.7582465277777778,
520
+ "0.9": 0.7020399305555555
521
+ },
522
+ "matched_frac_signed_band": 0.7557237413194444,
523
+ "matched_frac_signed_band_std": 0.009915146747727952,
524
+ "babel_mu1_abs_mean": 0.37611472606658936,
525
+ "babel_mu1_abs_std": 0.003267168590536138,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 652.3333333333334,
528
+ "0.6": 535.6666666666666,
529
+ "0.7": 410.6666666666667,
530
+ "0.8": 299.6666666666667,
531
+ "0.9": 180.66666666666666
532
+ },
533
+ "n_dup_abs_band": 356.6666666666667,
534
+ "n_dup_abs_band_std": 28.583576286625387,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 172.66666666666666,
537
+ "0.6": 138.33333333333334,
538
+ "0.7": 104.0,
539
+ "0.8": 72.33333333333333,
540
+ "0.9": 53.333333333333336
541
+ },
542
+ "n_dup_clusters_abs_band": 92.0,
543
+ "n_dup_clusters_abs_band_std": 2.462214450449026,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 80.33333333333333,
546
+ "0.6": 60.666666666666664,
547
+ "0.7": 45.0,
548
+ "0.8": 31.333333333333332,
549
+ "0.9": 11.666666666666666
550
+ },
551
+ "largest_dup_abs_band": 37.166666666666664,
552
+ "largest_dup_abs_band_std": 5.346338310781814,
553
+ "dcs_abs_mean": 0.8733628789583842,
554
+ "dcs_abs_std": 0.014599923677861167,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.7535627993994342,
557
+ "0.6": 0.7746064987214697,
558
+ "0.7": 0.7966051870963485,
559
+ "0.8": 0.8162697719267777,
560
+ "0.9": 0.8412849232069372
561
+ },
562
+ "dedup_dcs_abs_band": 0.8071915952378833,
563
+ "dedup_dcs_abs_band_std": 0.002126596326879785,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.8930074373881022,
566
+ "0.6": 0.889689306418101,
567
+ "0.7": 0.8849606911341349,
568
+ "0.8": 0.8815199335416158,
569
+ "0.9": 0.877691388130188
570
+ },
571
+ "matched_cos_abs_band": 0.8834653298060099,
572
+ "matched_cos_abs_band_std": 0.005914915546760715,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.8438585069444445,
575
+ "0.6": 0.8706597222222223,
576
+ "0.7": 0.9001736111111112,
577
+ "0.8": 0.925998263888889,
578
+ "0.9": 0.9585503472222223
579
+ },
580
+ "unique_frac_abs_band": 0.9138454861111112,
581
+ "unique_frac_abs_band_std": 0.00432777723906263,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.7771267361111112,
584
+ "0.6": 0.7847222222222222,
585
+ "0.7": 0.7791883680555555,
586
+ "0.8": 0.7578125,
587
+ "0.9": 0.7010633680555557
588
+ },
589
+ "matched_frac_abs_band": 0.7556966145833334,
590
+ "matched_frac_abs_band_std": 0.01008819208703062,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
vision/dinov2-b_layer10/mp/x4/l0_32/seed1/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "k_iters": 32,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 32,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 307,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 771000,
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "l0": 94.26257976653696,
20
+ "recon_r2": 0.975821934346816,
21
+ "l1_per_token": 160.12352198927357
22
+ },
23
+ "seed1": {
24
+ "l0": 96.0824747081712,
25
+ "recon_r2": 0.9758979026055418,
26
+ "l1_per_token": 160.55607186960745
27
+ },
28
+ "seed2": {
29
+ "l0": 96.16374837872893,
30
+ "recon_r2": 0.9758934109302257,
31
+ "l1_per_token": 160.4021775571133
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.686996152798104,
38
+ "mcc_abs": 0.6870703592996316,
39
+ "act_iou": 0.25547694536790305,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9846564578689908,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8233672574692353,
44
+ "0.7": 0.8235882716727134,
45
+ "0.8": 0.8303783596636519,
46
+ "0.9": 0.8476126278762411
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8221728981196199,
50
+ "0.7": 0.822716947685233,
51
+ "0.8": 0.831733380016577,
52
+ "0.9": 0.8478334868152545
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.6846840780527694,
58
+ "mcc_abs": 0.6847622419526243,
59
+ "act_iou": 0.25246337038642785,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9844104815350083,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8227527574304134,
64
+ "0.7": 0.8178582045636273,
65
+ "0.8": 0.8274592190546871,
66
+ "0.9": 0.844338840962079
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8202325549355041,
70
+ "0.7": 0.8191295136242774,
71
+ "0.8": 0.8274661268153322,
72
+ "0.9": 0.8450663089562113
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.6946268447282802,
78
+ "mcc_abs": 0.6946844550087684,
79
+ "act_iou": 0.2577866682674802,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9846777056136713,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.822279493121979,
84
+ "0.7": 0.818058897473975,
85
+ "0.8": 0.8203782644254983,
86
+ "0.9": 0.8269563531697791
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.820086935363402,
90
+ "0.7": 0.8187139526860329,
91
+ "0.8": 0.819773872370219,
92
+ "0.9": 0.8279184476094364
93
+ }
94
+ }
95
+ },
96
+ "summary": {
97
+ "n_seeds": 3,
98
+ "n_pairs": 3,
99
+ "report_band": [
100
+ 0.6,
101
+ 0.7,
102
+ 0.8,
103
+ 0.9
104
+ ],
105
+ "l0_mean": 95.50293428447902,
106
+ "l0_std": 1.0749469056730019,
107
+ "recon_r2_mean": 0.975871082627528,
108
+ "recon_r2_std": 4.2622868386407093e-05,
109
+ "l1_per_token_mean": 160.3605904719981,
110
+ "l1_per_token_std": 0.2192531960553768,
111
+ "input_r2_mean": 0.9845815483392236,
112
+ "input_r2_std": 0.00014852863458536,
113
+ "mcc_signed_mean": 0.6887690251930513,
114
+ "mcc_signed_std": 0.005203072139369793,
115
+ "mcc_abs_mean": 0.6888390187536748,
116
+ "mcc_abs_std": 0.005192176343419167,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8227998360072094,
119
+ "0.7": 0.8198351245701052,
120
+ "0.8": 0.8260719477146125,
121
+ "0.9": 0.8396359406693664
122
+ },
123
+ "matched_mcc_signed_band": 0.8270857122403233,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8208307961395086,
126
+ "0.7": 0.8201868046651811,
127
+ "0.8": 0.8263244597340428,
128
+ "0.9": 0.8402727477936341
129
+ },
130
+ "matched_mcc_abs_band": 0.8269037020830917,
131
+ "act_iou_mean": 0.25524232800727037,
132
+ "act_iou_std": 0.002669393013783618,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/dict_metrics_step_final.json ADDED
@@ -0,0 +1,595 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 3,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 3072,
20
+ "d_in": 768,
21
+ "l0": null,
22
+ "l0_raw": null,
23
+ "recon_r2": null,
24
+ "l1_per_token": null,
25
+ "stats_source": "mp-excluded",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 0,
30
+ "dense_rate": 0.0,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 633,
34
+ "0.6": 505,
35
+ "0.7": 383,
36
+ "0.8": 264,
37
+ "0.9": 129
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 178,
41
+ "0.6": 140,
42
+ "0.7": 110,
43
+ "0.8": 71,
44
+ "0.9": 33
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 455,
48
+ "0.6": 365,
49
+ "0.7": 273,
50
+ "0.8": 193,
51
+ "0.9": 96
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 102,
55
+ "0.6": 56,
56
+ "0.7": 46,
57
+ "0.8": 27,
58
+ "0.9": 14
59
+ },
60
+ "babel_mu1_signed": 0.3663797080516815,
61
+ "n_dup_abs": {
62
+ "0.5": 644,
63
+ "0.6": 511,
64
+ "0.7": 389,
65
+ "0.8": 267,
66
+ "0.9": 130
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 181,
70
+ "0.6": 142,
71
+ "0.7": 112,
72
+ "0.8": 72,
73
+ "0.9": 34
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 463,
77
+ "0.6": 369,
78
+ "0.7": 277,
79
+ "0.8": 195,
80
+ "0.9": 96
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 106,
84
+ "0.6": 55,
85
+ "0.7": 51,
86
+ "0.8": 27,
87
+ "0.9": 11
88
+ },
89
+ "babel_mu1_abs": 0.375279039144516,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": null,
96
+ "l0_raw": null,
97
+ "recon_r2": null,
98
+ "l1_per_token": null,
99
+ "stats_source": "mp-excluded",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 0,
104
+ "dense_rate": 0.0,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 658,
108
+ "0.6": 541,
109
+ "0.7": 413,
110
+ "0.8": 303,
111
+ "0.9": 198
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 174,
115
+ "0.6": 140,
116
+ "0.7": 100,
117
+ "0.8": 65,
118
+ "0.9": 60
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 484,
122
+ "0.6": 401,
123
+ "0.7": 313,
124
+ "0.8": 238,
125
+ "0.9": 138
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 89,
129
+ "0.6": 48,
130
+ "0.7": 28,
131
+ "0.8": 22,
132
+ "0.9": 10
133
+ },
134
+ "babel_mu1_signed": 0.3705531060695648,
135
+ "n_dup_abs": {
136
+ "0.5": 671,
137
+ "0.6": 548,
138
+ "0.7": 414,
139
+ "0.8": 308,
140
+ "0.9": 200
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 179,
144
+ "0.6": 140,
145
+ "0.7": 97,
146
+ "0.8": 67,
147
+ "0.9": 61
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 492,
151
+ "0.6": 408,
152
+ "0.7": 317,
153
+ "0.8": 241,
154
+ "0.9": 139
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 73,
158
+ "0.6": 70,
159
+ "0.7": 28,
160
+ "0.8": 22,
161
+ "0.9": 10
162
+ },
163
+ "babel_mu1_abs": 0.3797185719013214,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": null,
170
+ "l0_raw": null,
171
+ "recon_r2": null,
172
+ "l1_per_token": null,
173
+ "stats_source": "mp-excluded",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 0,
178
+ "dense_rate": 0.0,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 639,
182
+ "0.6": 540,
183
+ "0.7": 423,
184
+ "0.8": 318,
185
+ "0.9": 208
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 156,
189
+ "0.6": 128,
190
+ "0.7": 99,
191
+ "0.8": 74,
192
+ "0.9": 63
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 483,
196
+ "0.6": 412,
197
+ "0.7": 324,
198
+ "0.8": 244,
199
+ "0.9": 145
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 111,
203
+ "0.6": 57,
204
+ "0.7": 56,
205
+ "0.8": 45,
206
+ "0.9": 14
207
+ },
208
+ "babel_mu1_signed": 0.3651764392852783,
209
+ "n_dup_abs": {
210
+ "0.5": 642,
211
+ "0.6": 548,
212
+ "0.7": 429,
213
+ "0.8": 324,
214
+ "0.9": 212
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 158,
218
+ "0.6": 133,
219
+ "0.7": 103,
220
+ "0.8": 78,
221
+ "0.9": 65
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 484,
225
+ "0.6": 415,
226
+ "0.7": 326,
227
+ "0.8": 246,
228
+ "0.9": 147
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 62,
232
+ "0.6": 57,
233
+ "0.7": 56,
234
+ "0.8": 45,
235
+ "0.9": 14
236
+ },
237
+ "babel_mu1_abs": 0.37334656715393066,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.863697350025177,
244
+ "dcs_abs": 0.8660604953765869,
245
+ "frac_sign_flipped": 0.033528645833333336,
246
+ "alive_source": "stats",
247
+ "frac_denominator": "d_sae",
248
+ "n_alive_a": 3072,
249
+ "n_alive_b": 3072,
250
+ "matched_cos_signed": {
251
+ "0.5": 0.8904803991317749,
252
+ "0.6": 0.8863230347633362,
253
+ "0.7": 0.8811215758323669,
254
+ "0.8": 0.8779223561286926,
255
+ "0.9": 0.8697978258132935
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.84716796875,
259
+ "0.6": 0.8753255208333334,
260
+ "0.7": 0.9046223958333334,
261
+ "0.8": 0.9298502604166666,
262
+ "0.9": 0.9619140625
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.754386470944155,
266
+ "0.6": 0.7758211720307979,
267
+ "0.7": 0.7970823109499179,
268
+ "0.8": 0.8163363314718783,
269
+ "0.9": 0.8366707601817325
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.7802734375,
273
+ "0.6": 0.7838541666666666,
274
+ "0.7": 0.7750651041666666,
275
+ "0.8": 0.7529296875,
276
+ "0.9": 0.6888020833333334
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.892687201499939,
280
+ "0.6": 0.8885923027992249,
281
+ "0.7": 0.8832919001579285,
282
+ "0.8": 0.8800186514854431,
283
+ "0.9": 0.8724757432937622
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.8445638020833334,
287
+ "0.6": 0.87353515625,
288
+ "0.7": 0.9033203125,
289
+ "0.8": 0.9290364583333334,
290
+ "0.9": 0.9617513020833334
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.7539312969699192,
294
+ "0.6": 0.7762166160682682,
295
+ "0.7": 0.7978955152793787,
296
+ "0.8": 0.8175694112433121,
297
+ "0.9": 0.8391046821488999
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.7802734375,
301
+ "0.6": 0.7848307291666666,
302
+ "0.7": 0.7760416666666666,
303
+ "0.8": 0.751953125,
304
+ "0.9": 0.6884765625
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.8620738983154297,
309
+ "dcs_abs": 0.8638548851013184,
310
+ "frac_sign_flipped": 0.028645833333333332,
311
+ "alive_source": "stats",
312
+ "frac_denominator": "d_sae",
313
+ "n_alive_a": 3072,
314
+ "n_alive_b": 3072,
315
+ "matched_cos_signed": {
316
+ "0.5": 0.8867824077606201,
317
+ "0.6": 0.8854392170906067,
318
+ "0.7": 0.8801518082618713,
319
+ "0.8": 0.8756371140480042,
320
+ "0.9": 0.8707349896430969
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.8473307291666666,
324
+ "0.6": 0.87353515625,
325
+ "0.7": 0.90283203125,
326
+ "0.8": 0.9288736979166666,
327
+ "0.9": 0.9607747395833334
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.7513979841799785,
331
+ "0.6": 0.7734622848511208,
332
+ "0.7": 0.7946292448614258,
333
+ "0.8": 0.8133562841588475,
334
+ "0.9": 0.8365801829204429
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.7734375,
338
+ "0.6": 0.7809244791666666,
339
+ "0.7": 0.7770182291666666,
340
+ "0.8": 0.7513020833333334,
341
+ "0.9": 0.6901041666666666
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.8874406218528748,
345
+ "0.6": 0.8861882090568542,
346
+ "0.7": 0.8809804916381836,
347
+ "0.8": 0.8767107129096985,
348
+ "0.9": 0.8725345730781555
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.8458658854166666,
352
+ "0.6": 0.8723958333333334,
353
+ "0.7": 0.90185546875,
354
+ "0.8": 0.92822265625,
355
+ "0.9": 0.96044921875
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.7506557473582991,
359
+ "0.6": 0.7731069011303286,
360
+ "0.7": 0.7945170742459595,
361
+ "0.8": 0.8137827466998715,
362
+ "0.9": 0.8380251490452792
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.7737630208333334,
366
+ "0.6": 0.7822265625,
367
+ "0.7": 0.7760416666666666,
368
+ "0.8": 0.7506510416666666,
369
+ "0.9": 0.6888020833333334
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8889439702033997,
374
+ "dcs_abs": 0.8901732563972473,
375
+ "frac_sign_flipped": 0.0185546875,
376
+ "alive_source": "stats",
377
+ "frac_denominator": "d_sae",
378
+ "n_alive_a": 3072,
379
+ "n_alive_b": 3072,
380
+ "matched_cos_signed": {
381
+ "0.5": 0.8958790898323059,
382
+ "0.6": 0.893430769443512,
383
+ "0.7": 0.889135479927063,
384
+ "0.8": 0.8864941000938416,
385
+ "0.9": 0.8865686058998108
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.8426106770833334,
389
+ "0.6": 0.86767578125,
390
+ "0.7": 0.8963216145833334,
391
+ "0.8": 0.9215494791666666,
392
+ "0.9": 0.9539388020833334
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.7548772864683997,
396
+ "0.6": 0.7752082408696879,
397
+ "0.7": 0.7969513489515521,
398
+ "0.8": 0.8169481762258025,
399
+ "0.9": 0.8457321938767564
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.7760416666666666,
403
+ "0.6": 0.7858072916666666,
404
+ "0.7": 0.78515625,
405
+ "0.8": 0.7705078125,
406
+ "0.9": 0.7272135416666666
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.8988944888114929,
410
+ "0.6": 0.8942874073982239,
411
+ "0.7": 0.8906096816062927,
412
+ "0.8": 0.8878304362297058,
413
+ "0.9": 0.8880638480186462
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.8411458333333334,
417
+ "0.6": 0.8660481770833334,
418
+ "0.7": 0.8953450520833334,
419
+ "0.8": 0.9207356770833334,
420
+ "0.9": 0.9534505208333334
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.7561013538700839,
424
+ "0.6": 0.7744959789658121,
425
+ "0.7": 0.7974029717637071,
426
+ "0.8": 0.8174571578371495,
427
+ "0.9": 0.8467249384266324
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.77734375,
431
+ "0.6": 0.787109375,
432
+ "0.7": 0.7854817708333334,
433
+ "0.8": 0.7708333333333334,
434
+ "0.9": 0.7259114583333334
435
+ }
436
+ }
437
+ },
438
+ "summary": {
439
+ "n_seeds": 3,
440
+ "n_pairs": 3,
441
+ "report_band": [
442
+ 0.6,
443
+ 0.7,
444
+ 0.8,
445
+ 0.9
446
+ ],
447
+ "n_dead_mean": 0,
448
+ "n_dead_std": 0.0,
449
+ "n_dense_mean": 0,
450
+ "n_dense_std": 0.0,
451
+ "l0_mean": null,
452
+ "l0_std": null,
453
+ "recon_r2_mean": null,
454
+ "recon_r2_std": null,
455
+ "l1_per_token_mean": null,
456
+ "l1_per_token_std": null,
457
+ "babel_mu1_signed_mean": 0.36736975113550824,
458
+ "babel_mu1_signed_std": 0.0028217504156440275,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 643.3333333333334,
461
+ "0.6": 528.6666666666666,
462
+ "0.7": 406.3333333333333,
463
+ "0.8": 295.0,
464
+ "0.9": 178.33333333333334
465
+ },
466
+ "n_dup_signed_band": 352.0833333333333,
467
+ "n_dup_signed_band_std": 27.894145144336893,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 169.33333333333334,
470
+ "0.6": 136.0,
471
+ "0.7": 103.0,
472
+ "0.8": 70.0,
473
+ "0.9": 52.0
474
+ },
475
+ "n_dup_clusters_signed_band": 90.25,
476
+ "n_dup_clusters_signed_band_std": 1.5206906325745548,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 100.66666666666667,
479
+ "0.6": 53.666666666666664,
480
+ "0.7": 43.333333333333336,
481
+ "0.8": 31.333333333333332,
482
+ "0.9": 12.666666666666666
483
+ },
484
+ "largest_dup_signed_band": 35.25,
485
+ "largest_dup_signed_band_std": 8.011710179481033,
486
+ "dcs_signed_mean": 0.8715717395146688,
487
+ "dcs_signed_std": 0.015066675087434934,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.7535539138641777,
490
+ "0.6": 0.7748305659172022,
491
+ "0.7": 0.7962209682542986,
492
+ "0.8": 0.8155469306188428,
493
+ "0.9": 0.839661045659644
494
+ },
495
+ "dedup_dcs_signed_band": 0.8065648776124968,
496
+ "dedup_dcs_signed_band_std": 0.002102852872293367,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.8910472989082336,
499
+ "0.6": 0.8883976737658182,
500
+ "0.7": 0.8834696213404337,
501
+ "0.8": 0.8800178567568461,
502
+ "0.9": 0.8757004737854004
503
+ },
504
+ "matched_cos_signed_band": 0.8818964064121246,
505
+ "matched_cos_signed_band_std": 0.006084734579468042,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.845703125,
508
+ "0.6": 0.8721788194444445,
509
+ "0.7": 0.9012586805555557,
510
+ "0.8": 0.9267578125,
511
+ "0.9": 0.9588758680555557
512
+ },
513
+ "unique_frac_signed_band": 0.914767795138889,
514
+ "unique_frac_signed_band_std": 0.004299758816805034,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.7765842013888888,
517
+ "0.6": 0.7835286458333334,
518
+ "0.7": 0.779079861111111,
519
+ "0.8": 0.7582465277777778,
520
+ "0.9": 0.7020399305555555
521
+ },
522
+ "matched_frac_signed_band": 0.7557237413194444,
523
+ "matched_frac_signed_band_std": 0.009915146747727952,
524
+ "babel_mu1_abs_mean": 0.37611472606658936,
525
+ "babel_mu1_abs_std": 0.003267168590536138,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 652.3333333333334,
528
+ "0.6": 535.6666666666666,
529
+ "0.7": 410.6666666666667,
530
+ "0.8": 299.6666666666667,
531
+ "0.9": 180.66666666666666
532
+ },
533
+ "n_dup_abs_band": 356.6666666666667,
534
+ "n_dup_abs_band_std": 28.583576286625387,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 172.66666666666666,
537
+ "0.6": 138.33333333333334,
538
+ "0.7": 104.0,
539
+ "0.8": 72.33333333333333,
540
+ "0.9": 53.333333333333336
541
+ },
542
+ "n_dup_clusters_abs_band": 92.0,
543
+ "n_dup_clusters_abs_band_std": 2.462214450449026,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 80.33333333333333,
546
+ "0.6": 60.666666666666664,
547
+ "0.7": 45.0,
548
+ "0.8": 31.333333333333332,
549
+ "0.9": 11.666666666666666
550
+ },
551
+ "largest_dup_abs_band": 37.166666666666664,
552
+ "largest_dup_abs_band_std": 5.346338310781814,
553
+ "dcs_abs_mean": 0.8733628789583842,
554
+ "dcs_abs_std": 0.014599923677861167,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.7535627993994342,
557
+ "0.6": 0.7746064987214697,
558
+ "0.7": 0.7966051870963485,
559
+ "0.8": 0.8162697719267777,
560
+ "0.9": 0.8412849232069372
561
+ },
562
+ "dedup_dcs_abs_band": 0.8071915952378833,
563
+ "dedup_dcs_abs_band_std": 0.002126596326879785,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.8930074373881022,
566
+ "0.6": 0.889689306418101,
567
+ "0.7": 0.8849606911341349,
568
+ "0.8": 0.8815199335416158,
569
+ "0.9": 0.877691388130188
570
+ },
571
+ "matched_cos_abs_band": 0.8834653298060099,
572
+ "matched_cos_abs_band_std": 0.005914915546760715,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.8438585069444445,
575
+ "0.6": 0.8706597222222223,
576
+ "0.7": 0.9001736111111112,
577
+ "0.8": 0.925998263888889,
578
+ "0.9": 0.9585503472222223
579
+ },
580
+ "unique_frac_abs_band": 0.9138454861111112,
581
+ "unique_frac_abs_band_std": 0.00432777723906263,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.7771267361111112,
584
+ "0.6": 0.7847222222222222,
585
+ "0.7": 0.7791883680555555,
586
+ "0.8": 0.7578125,
587
+ "0.9": 0.7010633680555557
588
+ },
589
+ "matched_frac_abs_band": 0.7556966145833334,
590
+ "matched_frac_abs_band_std": 0.01008819208703062,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed0/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "k_iters": 32,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 32,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 307,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1779,
3
+ "r2_score": 0.9518,
4
+ "l0_per_token_0": 0.0077,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 93.5173
9
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed1/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "k_iters": 32,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 32,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 307,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1793,
3
+ "r2_score": 0.9515,
4
+ "l0_per_token_0": 0.0076,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 92.3942
9
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed2/config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "mp",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "k_iters": 32,
9
+ "mp_nonneg": true,
10
+ "mp_dropout": 0.0,
11
+ "inference_mode": "adaptive",
12
+ "inference_iters": 32,
13
+ "residual_thr": 0.05,
14
+ "max_iters": 307,
15
+ "patience": 2,
16
+ "class_type": "MPSAE",
17
+ "module_name": "consdl.models.mp"
18
+ }
vision/dinov2-b_layer10/mp/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.179,
3
+ "r2_score": 0.9515,
4
+ "l0_per_token_0": 0.0077,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0,
7
+ "live_latent_frac": 1.0,
8
+ "l1_per_token": 92.3052
9
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed0/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1393,
3
+ "r2_score": 0.9623,
4
+ "l0_per_token_0": 0.0101,
5
+ "l0_per_token_0.001": 0.0101,
6
+ "l0_per_token_0.01": 0.0098,
7
+ "l0_per_token_0.05": 0.0084,
8
+ "l0_per_token_0.1": 0.0071,
9
+ "l0_per_token_0.5": 0.0025,
10
+ "l0_per_token_1.0": 0.0012,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0001,
13
+ "live_latent_frac": 0.996,
14
+ "l1_per_token": 65.3288
15
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed1/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1388,
3
+ "r2_score": 0.9624,
4
+ "l0_per_token_0": 0.0102,
5
+ "l0_per_token_0.001": 0.0102,
6
+ "l0_per_token_0.01": 0.0098,
7
+ "l0_per_token_0.05": 0.0085,
8
+ "l0_per_token_0.1": 0.0071,
9
+ "l0_per_token_0.5": 0.0025,
10
+ "l0_per_token_1.0": 0.0012,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0001,
13
+ "live_latent_frac": 0.9961,
14
+ "l1_per_token": 65.4799
15
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed2/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.1395,
3
+ "r2_score": 0.9639,
4
+ "l0_per_token_0": 0.0102,
5
+ "dead_rate": 0.0,
6
+ "dense_rate_0.2": 0.0002,
7
+ "live_latent_frac": 0.9976,
8
+ "l1_per_token": 65.1052
9
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed0/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.2285,
3
+ "r2_score": 0.9381,
4
+ "l0_per_token_0": 0.0027,
5
+ "l0_per_token_0.001": 0.0027,
6
+ "l0_per_token_0.01": 0.0027,
7
+ "l0_per_token_0.05": 0.0025,
8
+ "l0_per_token_0.1": 0.0024,
9
+ "l0_per_token_0.5": 0.0015,
10
+ "l0_per_token_1.0": 0.001,
11
+ "dead_rate": 0.0046,
12
+ "dense_rate_0.2": 0.0001,
13
+ "live_latent_frac": 0.7634,
14
+ "l1_per_token": 44.3201
15
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed1/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.228,
3
+ "r2_score": 0.9382,
4
+ "l0_per_token_0": 0.0027,
5
+ "l0_per_token_0.001": 0.0027,
6
+ "l0_per_token_0.01": 0.0027,
7
+ "l0_per_token_0.05": 0.0025,
8
+ "l0_per_token_0.1": 0.0024,
9
+ "l0_per_token_0.5": 0.0015,
10
+ "l0_per_token_1.0": 0.001,
11
+ "dead_rate": 0.0098,
12
+ "dense_rate_0.2": 0.0001,
13
+ "live_latent_frac": 0.7194,
14
+ "l1_per_token": 44.3249
15
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed2/config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "relu",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "class_type": "VanillaSAE",
9
+ "module_name": "fastsae.models.sae"
10
+ }
vision/dinov2-b_layer10/relu/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.2272,
3
+ "r2_score": 0.9385,
4
+ "l0_per_token_0": 0.0027,
5
+ "l0_per_token_0.001": 0.0027,
6
+ "l0_per_token_0.01": 0.0027,
7
+ "l0_per_token_0.05": 0.0025,
8
+ "l0_per_token_0.1": 0.0024,
9
+ "l0_per_token_0.5": 0.0015,
10
+ "l0_per_token_1.0": 0.001,
11
+ "dead_rate": 0.0091,
12
+ "dense_rate_0.2": 0.0001,
13
+ "live_latent_frac": 0.7134,
14
+ "l1_per_token": 44.2982
15
+ }
vision/dinov2-b_layer10/relu/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "l2_loss": 0.2942,
3
+ "r2_score": 0.9203,
4
+ "l0_per_token_0": 0.0154,
5
+ "l0_per_token_0.001": 0.0152,
6
+ "l0_per_token_0.01": 0.014,
7
+ "l0_per_token_0.05": 0.0101,
8
+ "l0_per_token_0.1": 0.0071,
9
+ "l0_per_token_0.5": 0.0019,
10
+ "l0_per_token_1.0": 0.0009,
11
+ "dead_rate": 0.0,
12
+ "dense_rate_0.2": 0.0007,
13
+ "live_latent_frac": 1.0,
14
+ "l1_per_token": 55.4259
15
+ }