jjho-choi commited on
Commit
daf82d0
·
verified ·
1 Parent(s): c69db0b

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. lm/gpt2-small_layer8/consdl/x16/l0_128/code_metrics_step_final.json +136 -0
  2. lm/gpt2-small_layer8/consdl/x16/l0_128/dict_metrics_step_final.json +550 -0
  3. lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/config.json +24 -0
  4. lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  5. lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  6. lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/config.json +24 -0
  7. lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  8. lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  9. lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/config.json +24 -0
  10. lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  11. lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json +26 -0
  12. lm/gpt2-small_layer8/consdl/x16/l0_32/code_metrics_step_final.json +136 -0
  13. lm/gpt2-small_layer8/consdl/x16/l0_32/dict_metrics_step_final.json +550 -0
  14. lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/config.json +24 -0
  15. lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  16. lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  17. lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/config.json +24 -0
  18. lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  19. lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  20. lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/config.json +24 -0
  21. lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  22. lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/stats.fineweb.split_train.seed_no/meta.json +26 -0
  23. lm/gpt2-small_layer8/consdl/x16/l0_64/code_metrics_step_final.json +136 -0
  24. lm/gpt2-small_layer8/consdl/x16/l0_64/dict_metrics_step_final.json +550 -0
  25. lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/config.json +24 -0
  26. lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  27. lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  28. lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/config.json +24 -0
  29. lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  30. lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  31. lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/config.json +24 -0
  32. lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  33. lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/stats.fineweb.split_train.seed_no/meta.json +26 -0
  34. lm/gpt2-small_layer8/consdl/x4/l0_128/code_metrics_step_final.json +136 -0
  35. lm/gpt2-small_layer8/consdl/x4/l0_128/dict_metrics_step_final.json +550 -0
  36. lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/config.json +24 -0
  37. lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  38. lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  39. lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/config.json +24 -0
  40. lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  41. lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  42. lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/config.json +24 -0
  43. lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  44. lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json +26 -0
  45. lm/gpt2-small_layer8/consdl/x4/l0_32/code_metrics_step_final.json +136 -0
  46. lm/gpt2-small_layer8/consdl/x4/l0_32/dict_metrics_step_final.json +595 -0
  47. lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/config.json +24 -0
  48. lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  49. lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  50. lm/gpt2-small_layer8/consdl/x4/l0_32/seed1/config.json +24 -0
lm/gpt2-small_layer8/consdl/x16/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 600000,
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": 126.42603333333334,
20
+ "recon_r2": 0.943795649182259,
21
+ "l1_per_token": 289.4212340388216
22
+ },
23
+ "seed1": {
24
+ "l0": 125.811965,
25
+ "recon_r2": 0.9438612213988956,
26
+ "l1_per_token": 289.53429910061254
27
+ },
28
+ "seed2": {
29
+ "l0": 126.05112333333334,
30
+ "recon_r2": 0.9438580555624618,
31
+ "l1_per_token": 289.6181956334526
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7859704474411874,
38
+ "mcc_abs": 0.7859711613635518,
39
+ "act_iou": 0.45132704906101173,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9913803291545396,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8342782215099661,
44
+ "0.7": 0.8634063066238307,
45
+ "0.8": 0.8990084994065084,
46
+ "0.9": 0.9396039340469318
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8335566129871208,
50
+ "0.7": 0.8630353633737552,
51
+ "0.8": 0.8988650385049328,
52
+ "0.9": 0.9395716929205603
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.7828689400311742,
58
+ "mcc_abs": 0.7828697764077752,
59
+ "act_iou": 0.4478816509665113,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.991334920205118,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8309814935810688,
64
+ "0.7": 0.8613655575710586,
65
+ "0.8": 0.8984125478906685,
66
+ "0.9": 0.9399642265869411
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8310063654716346,
70
+ "0.7": 0.8610575981792729,
71
+ "0.8": 0.8984265823137941,
72
+ "0.9": 0.9399718561318022
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.783658469940665,
78
+ "mcc_abs": 0.7836587034705734,
79
+ "act_iou": 0.4499608370893986,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.9913863229725707,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8351966666901895,
84
+ "0.7": 0.8621449988701696,
85
+ "0.8": 0.8981497940397523,
86
+ "0.9": 0.939461465954832
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8341974655594866,
90
+ "0.7": 0.8619405907398374,
91
+ "0.8": 0.8981406120242229,
92
+ "0.9": 0.939461465954832
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": 126.09637388888889,
106
+ "l0_std": 0.30952494097287203,
107
+ "recon_r2_mean": 0.9438383087145388,
108
+ "recon_r2_std": 3.6978134090712164e-05,
109
+ "l1_per_token_mean": 289.5245762576289,
110
+ "l1_per_token_std": 0.09884011177990036,
111
+ "input_r2_mean": 0.9913671907774094,
112
+ "input_r2_std": 2.810736274817572e-05,
113
+ "mcc_signed_mean": 0.7841659524710088,
114
+ "mcc_signed_std": 0.0016118284469957588,
115
+ "mcc_abs_mean": 0.7841665470806335,
116
+ "mcc_abs_std": 0.00161185483533366,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8334854605937414,
119
+ "0.7": 0.8623056210216863,
120
+ "0.8": 0.8985236137789764,
121
+ "0.9": 0.939676542196235
122
+ },
123
+ "matched_mcc_signed_band": 0.8834978093976598,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8329201480060807,
126
+ "0.7": 0.8620111840976218,
127
+ "0.8": 0.89847741094765,
128
+ "0.9": 0.9396683383357316
129
+ },
130
+ "matched_mcc_abs_band": 0.883269270346771,
131
+ "act_iou_mean": 0.4497231790389739,
132
+ "act_iou_std": 0.0017349504370717912,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 126,
22
+ "l0_raw": 126.4695044921875,
23
+ "recon_r2": 0.9436933468059865,
24
+ "l1_per_token": 289.3836344397387,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 19,
30
+ "dense_rate": 0.0015462239583333333,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 3965,
34
+ "0.6": 19,
35
+ "0.7": 3,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 1804,
41
+ "0.6": 7,
42
+ "0.7": 1,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 8,
48
+ "0.6": 4,
49
+ "0.7": 3,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.4519537389278412,
54
+ "n_dup_abs": {
55
+ "0.5": 4019,
56
+ "0.6": 117,
57
+ "0.7": 19,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 1831,
63
+ "0.6": 56,
64
+ "0.7": 9,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 8,
70
+ "0.6": 4,
71
+ "0.7": 3,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.45511364936828613
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 126,
81
+ "l0_raw": 125.89824155273438,
82
+ "recon_r2": 0.9437663728355269,
83
+ "l1_per_token": 289.51439305719344,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 19,
89
+ "dense_rate": 0.0015462239583333333,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 3819,
93
+ "0.6": 14,
94
+ "0.7": 2,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 1739,
100
+ "0.6": 4,
101
+ "0.7": 1,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 7,
107
+ "0.6": 7,
108
+ "0.7": 2,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.4506072700023651,
113
+ "n_dup_abs": {
114
+ "0.5": 3913,
115
+ "0.6": 124,
116
+ "0.7": 10,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 1787,
122
+ "0.6": 59,
123
+ "0.7": 5,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 7,
129
+ "0.6": 7,
130
+ "0.7": 2,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.45388028025627136
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 126,
140
+ "l0_raw": 126.127159765625,
141
+ "recon_r2": 0.9437536036438203,
142
+ "l1_per_token": 289.6085860579297,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 20,
148
+ "dense_rate": 0.0016276041666666667,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 3873,
152
+ "0.6": 21,
153
+ "0.7": 2,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 1766,
159
+ "0.6": 8,
160
+ "0.7": 1,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 8,
166
+ "0.6": 5,
167
+ "0.7": 2,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.45112404227256775,
172
+ "n_dup_abs": {
173
+ "0.5": 3969,
174
+ "0.6": 125,
175
+ "0.7": 12,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 1813,
181
+ "0.6": 60,
182
+ "0.7": 6,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 8,
188
+ "0.6": 5,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.4543568193912506
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.8446399569511414,
199
+ "dcs_abs": 0.8448325991630554,
200
+ "frac_sign_flipped": 0.00390625,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 12288,
204
+ "n_alive_b": 12288,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.8421996831893921,
207
+ "0.6": 0.8446352481842041,
208
+ "0.7": 0.8446696400642395,
209
+ "0.8": 0.8446400165557861,
210
+ "0.9": 0.8446400165557861
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.8274332682291666,
214
+ "0.6": 0.9991048177083334,
215
+ "0.7": 0.9998779296875,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.6968640363629675,
221
+ "0.6": 0.8438791456671122,
222
+ "0.7": 0.8445665309773176,
223
+ "0.8": 0.8446400165557861,
224
+ "0.9": 0.8446400165557861
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.772705078125,
228
+ "0.6": 0.9130045572916666,
229
+ "0.7": 0.8390299479166666,
230
+ "0.8": 0.719970703125,
231
+ "0.9": 0.5185546875
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.8424644470214844,
235
+ "0.6": 0.8453661203384399,
236
+ "0.7": 0.8450070023536682,
237
+ "0.8": 0.8448325991630554,
238
+ "0.9": 0.8448325991630554
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.824462890625,
242
+ "0.6": 0.994873046875,
243
+ "0.7": 0.9993896484375,
244
+ "0.8": 1.0,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.6945806732401252,
249
+ "0.6": 0.8410319678660017,
250
+ "0.7": 0.8444912510094582,
251
+ "0.8": 0.8448325991630554,
252
+ "0.9": 0.8448325991630554
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.7703450520833334,
256
+ "0.6": 0.908935546875,
257
+ "0.7": 0.8392740885416666,
258
+ "0.8": 0.7202962239583334,
259
+ "0.9": 0.518798828125
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.8424331545829773,
264
+ "dcs_abs": 0.8425608277320862,
265
+ "frac_sign_flipped": 0.0032552083333333335,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 12288,
269
+ "n_alive_b": 12288,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.8382132649421692,
272
+ "0.6": 0.8423312306404114,
273
+ "0.7": 0.8424346446990967,
274
+ "0.8": 0.8424331545829773,
275
+ "0.9": 0.8424331545829773
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.8263346354166666,
279
+ "0.6": 0.9989827473958334,
280
+ "0.7": 0.9998779296875,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.6926446526874012,
286
+ "0.6": 0.8414743670024716,
287
+ "0.7": 0.8423318084387574,
288
+ "0.8": 0.8424331545829773,
289
+ "0.9": 0.8424331545829773
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.77197265625,
293
+ "0.6": 0.9117024739583334,
294
+ "0.7": 0.83349609375,
295
+ "0.8": 0.7119140625,
296
+ "0.9": 0.5091959635416666
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.8384155631065369,
300
+ "0.6": 0.8430929183959961,
301
+ "0.7": 0.8426738381385803,
302
+ "0.8": 0.8425607681274414,
303
+ "0.9": 0.8425607681274414
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.8232421875,
307
+ "0.6": 0.994873046875,
308
+ "0.7": 0.9993489583333334,
309
+ "0.8": 1.0,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.6902190622058697,
314
+ "0.6": 0.8387704205233604,
315
+ "0.7": 0.8421252223585423,
316
+ "0.8": 0.8425607681274414,
317
+ "0.9": 0.8425607681274414
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.7692057291666666,
321
+ "0.6": 0.9073079427083334,
322
+ "0.7": 0.8330891927083334,
323
+ "0.8": 0.711669921875,
324
+ "0.9": 0.5091145833333334
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.843366801738739,
329
+ "dcs_abs": 0.8435181975364685,
330
+ "frac_sign_flipped": 0.0035807291666666665,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 12288,
334
+ "n_alive_b": 12288,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.8385906219482422,
337
+ "0.6": 0.8433858156204224,
338
+ "0.7": 0.8433728814125061,
339
+ "0.8": 0.843366801738739,
340
+ "0.9": 0.843366801738739
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.8296305338541666,
344
+ "0.6": 0.9990641276041666,
345
+ "0.7": 0.9999186197916666,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.6957203853720177,
351
+ "0.6": 0.8425965141165458,
352
+ "0.7": 0.8433042475517141,
353
+ "0.8": 0.843366801738739,
354
+ "0.9": 0.843366801738739
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.773681640625,
358
+ "0.6": 0.9065755208333334,
359
+ "0.7": 0.8380533854166666,
360
+ "0.8": 0.7194010416666666,
361
+ "0.9": 0.5166829427083334
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.8398551940917969,
365
+ "0.6": 0.8440368175506592,
366
+ "0.7": 0.8436010479927063,
367
+ "0.8": 0.8435182571411133,
368
+ "0.9": 0.8435182571411133
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.8257649739583334,
372
+ "0.6": 0.9947102864583334,
373
+ "0.7": 0.9995524088541666,
374
+ "0.8": 1.0,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.6935230024779837,
379
+ "0.6": 0.8395721045671962,
380
+ "0.7": 0.843223459633009,
381
+ "0.8": 0.8435182571411133,
382
+ "0.9": 0.8435182571411133
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.7705891927083334,
386
+ "0.6": 0.904541015625,
387
+ "0.7": 0.8382161458333334,
388
+ "0.8": 0.7191569010416666,
389
+ "0.9": 0.5166829427083334
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 19,
405
+ "n_dense_std": 0.5773502691896258,
406
+ "l0_mean": 126,
407
+ "l0_std": 0.287502117670696,
408
+ "recon_r2_mean": 0.9437377744284445,
409
+ "recon_r2_std": 3.900158063678982e-05,
410
+ "l1_per_token_mean": 289.5022045182873,
411
+ "l1_per_token_std": 0.1129700313896754,
412
+ "babel_mu1_signed_mean": 0.4512283504009247,
413
+ "babel_mu1_signed_std": 0.0006792678271853955,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 3885.6666666666665,
416
+ "0.6": 18.0,
417
+ "0.7": 2.3333333333333335,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 5.083333333333333,
422
+ "n_dup_signed_band_std": 0.9464847243000456,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 1769.6666666666667,
425
+ "0.6": 6.333333333333333,
426
+ "0.7": 1.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 1.8333333333333333,
431
+ "n_dup_clusters_signed_band_std": 0.5204164998665332,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 7.666666666666667,
434
+ "0.6": 5.333333333333333,
435
+ "0.7": 2.3333333333333335,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 2.4166666666666665,
440
+ "largest_dup_signed_band_std": 0.2886751345948129,
441
+ "dcs_signed_mean": 0.8434799710909525,
442
+ "dcs_signed_std": 0.0011077453000324367,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.6950763581407955,
445
+ "0.6": 0.8426500089287098,
446
+ "0.7": 0.8434008623225964,
447
+ "0.8": 0.8434799909591675,
448
+ "0.9": 0.8434799909591675
449
+ },
450
+ "dedup_dcs_signed_band": 0.8432527132924102,
451
+ "dedup_dcs_signed_band_std": 0.001134584968802106,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.8396678566932678,
454
+ "0.6": 0.8434507648150126,
455
+ "0.7": 0.8434923887252808,
456
+ "0.8": 0.8434799909591675,
457
+ "0.9": 0.8434799909591675
458
+ },
459
+ "matched_cos_signed_band": 0.843475783864657,
460
+ "matched_cos_signed_band_std": 0.0011226214664995392,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.8277994791666666,
463
+ "0.6": 0.9990505642361112,
464
+ "0.7": 0.9998914930555555,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 0.9997355143229166,
469
+ "unique_frac_signed_band_std": 1.7619331945452135e-05,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.7727864583333334,
472
+ "0.6": 0.9104275173611112,
473
+ "0.7": 0.8368598090277777,
474
+ "0.8": 0.7170952690972222,
475
+ "0.9": 0.5148111979166666
476
+ },
477
+ "matched_frac_signed_band": 0.7447984483506944,
478
+ "matched_frac_signed_band_std": 0.003049202308208975,
479
+ "babel_mu1_abs_mean": 0.45445024967193604,
480
+ "babel_mu1_abs_std": 0.0006219700592711387,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 3967.0,
483
+ "0.6": 122.0,
484
+ "0.7": 13.666666666666666,
485
+ "0.8": 0.0,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 33.916666666666664,
489
+ "n_dup_abs_band_std": 0.38188130791298663,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 1810.3333333333333,
492
+ "0.6": 58.333333333333336,
493
+ "0.7": 6.666666666666667,
494
+ "0.8": 0.0,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 16.25,
498
+ "n_dup_clusters_abs_band_std": 0.25,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 7.666666666666667,
501
+ "0.6": 5.333333333333333,
502
+ "0.7": 2.3333333333333335,
503
+ "0.8": 1.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 2.4166666666666665,
507
+ "largest_dup_abs_band_std": 0.2886751345948129,
508
+ "dcs_abs_mean": 0.84363720814387,
509
+ "dcs_abs_std": 0.0011405520602530892,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.6927742459746596,
512
+ "0.6": 0.8397914976521861,
513
+ "0.7": 0.8432799776670032,
514
+ "0.8": 0.84363720814387,
515
+ "0.9": 0.84363720814387
516
+ },
517
+ "dedup_dcs_abs_band": 0.8425864729017323,
518
+ "dedup_dcs_abs_band_std": 0.0011517895005007416,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.8402450680732727,
521
+ "0.6": 0.8441652854283651,
522
+ "0.7": 0.843760629494985,
523
+ "0.8": 0.84363720814387,
524
+ "0.9": 0.84363720814387
525
+ },
526
+ "matched_cos_abs_band": 0.8438000828027725,
527
+ "matched_cos_abs_band_std": 0.001149408075698793,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.8244900173611112,
530
+ "0.6": 0.9948187934027778,
531
+ "0.7": 0.9994303385416666,
532
+ "0.8": 1.0,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.9985622829861112,
536
+ "unique_frac_abs_band_std": 5.873110648473362e-06,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.7700466579861112,
539
+ "0.6": 0.9069281684027778,
540
+ "0.7": 0.8368598090277778,
541
+ "0.8": 0.717041015625,
542
+ "0.9": 0.514865451388889
543
+ },
544
+ "matched_frac_abs_band": 0.743923611111111,
545
+ "matched_frac_abs_band_std": 0.0033253011024323635,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 4.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 126.4695044921875,
5
+ "l0_per_token_0.001": 126.46924731445313,
6
+ "l0_per_token_0.01": 126.46831743164063,
7
+ "l0_per_token_0.05": 126.46364501953126,
8
+ "l0_per_token_0.1": 126.454286328125,
9
+ "l0_per_token_0.5": 111.81361259765625,
10
+ "l0_per_token_1.0": 79.31354155273438,
11
+ "l1_per_token": 289.3836344397387,
12
+ "r2_score": 0.9436933468059865,
13
+ "l2_loss": 1.3205280686668774
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 4.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 125.89824155273438,
5
+ "l0_per_token_0.001": 125.89810434570313,
6
+ "l0_per_token_0.01": 125.89738203125,
7
+ "l0_per_token_0.05": 125.89346733398438,
8
+ "l0_per_token_0.1": 125.88476430664062,
9
+ "l0_per_token_0.5": 111.54088544921875,
10
+ "l0_per_token_1.0": 79.13532709960937,
11
+ "l1_per_token": 289.51439305719344,
12
+ "r2_score": 0.9437663728355269,
13
+ "l2_loss": 1.3188154305276674
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 4.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 10,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 126.127159765625,
5
+ "l0_per_token_0.001": 126.12707016601563,
6
+ "l0_per_token_0.01": 126.12630732421874,
7
+ "l0_per_token_0.05": 126.12196870117188,
8
+ "l0_per_token_0.1": 126.11321782226563,
9
+ "l0_per_token_0.5": 111.70350498046875,
10
+ "l0_per_token_1.0": 79.29182919921875,
11
+ "l1_per_token": 289.6085860579297,
12
+ "r2_score": 0.9437536036438203,
13
+ "l2_loss": 1.3191148991535948
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 600000,
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": 34.11666833333334,
20
+ "recon_r2": 0.8986679004156218,
21
+ "l1_per_token": 186.22070988001735
22
+ },
23
+ "seed1": {
24
+ "l0": 34.10902166666666,
25
+ "recon_r2": 0.8987894297283623,
26
+ "l1_per_token": 185.72599150414578
27
+ },
28
+ "seed2": {
29
+ "l0": 34.240206666666666,
30
+ "recon_r2": 0.8987818586148901,
31
+ "l1_per_token": 186.73523437882213
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7809844014446874,
38
+ "mcc_abs": 0.780985660585697,
39
+ "act_iou": 0.5074479851575096,
40
+ "act_iou_n_pairs": 12265,
41
+ "input_r2": 0.9863562707505128,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8304022051494929,
44
+ "0.7": 0.857498462100996,
45
+ "0.8": 0.8958479882608926,
46
+ "0.9": 0.9414620631216186
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8308284407037294,
50
+ "0.7": 0.8573308349313363,
51
+ "0.8": 0.895670010459787,
52
+ "0.9": 0.9414558876475153
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.7779173925424955,
58
+ "mcc_abs": 0.7779183138998308,
59
+ "act_iou": 0.5054622423786899,
60
+ "act_iou_n_pairs": 12271,
61
+ "input_r2": 0.9863696984638582,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8251178698395323,
64
+ "0.7": 0.8545442564293723,
65
+ "0.8": 0.8941268562305976,
66
+ "0.9": 0.9405185795513432
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8252502187072939,
70
+ "0.7": 0.8545884774945348,
71
+ "0.8": 0.8943297597748052,
72
+ "0.9": 0.9404768194422161
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.7805839115016303,
78
+ "mcc_abs": 0.7805845164009844,
79
+ "act_iou": 0.5090701067602436,
80
+ "act_iou_n_pairs": 12268,
81
+ "input_r2": 0.9864737100906098,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8274534585829221,
84
+ "0.7": 0.8567122593646093,
85
+ "0.8": 0.8951611853298442,
86
+ "0.9": 0.9415611179795759
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8272814411540946,
90
+ "0.7": 0.8561547899705673,
91
+ "0.8": 0.8952639714169829,
92
+ "0.9": 0.9415610600249489
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": 34.155298888888886,
106
+ "l0_std": 0.0736316231190653,
107
+ "recon_r2_mean": 0.8987463962529582,
108
+ "recon_r2_std": 6.808471046528173e-05,
109
+ "l1_per_token_mean": 186.2273119209951,
110
+ "l1_per_token_std": 0.5046538271231552,
111
+ "input_r2_mean": 0.9863998931016603,
112
+ "input_r2_std": 6.427897606995591e-05,
113
+ "mcc_signed_mean": 0.779828568496271,
114
+ "mcc_signed_std": 0.0016671962072113684,
115
+ "mcc_abs_mean": 0.7798294969621707,
116
+ "mcc_abs_std": 0.0016672416404239093,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8276578445239825,
119
+ "0.7": 0.8562516592983259,
120
+ "0.8": 0.8950453432737783,
121
+ "0.9": 0.9411805868841793
122
+ },
123
+ "matched_mcc_signed_band": 0.8800338584950664,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8277867001883726,
126
+ "0.7": 0.8560247007988128,
127
+ "0.8": 0.8950879138838584,
128
+ "0.9": 0.9411645890382268
129
+ },
130
+ "matched_mcc_abs_band": 0.8800159759773176,
131
+ "act_iou_mean": 0.5073267780988143,
132
+ "act_iou_std": 0.001806983594897617,
133
+ "act_iou_n_pairs": 12268.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 34,
22
+ "l0_raw": 34.09826708984375,
23
+ "recon_r2": 0.8984899544332474,
24
+ "l1_per_token": 186.09178804915646,
25
+ "stats_source": "stats",
26
+ "n_dead": 32,
27
+ "dead_rate": 0.0026041666666666665,
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": 5875,
34
+ "0.6": 123,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 2667,
41
+ "0.6": 60,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 6,
48
+ "0.6": 3,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.48740267753601074,
54
+ "n_dup_abs": {
55
+ "0.5": 5894,
56
+ "0.6": 193,
57
+ "0.7": 12,
58
+ "0.8": 0,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 2671,
63
+ "0.6": 95,
64
+ "0.7": 6,
65
+ "0.8": 0,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 6,
70
+ "0.6": 3,
71
+ "0.7": 2,
72
+ "0.8": 1,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.4891930818557739
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 34,
81
+ "l0_raw": 34.087190966796875,
82
+ "recon_r2": 0.8986077841689825,
83
+ "l1_per_token": 185.59309232967692,
84
+ "stats_source": "stats",
85
+ "n_dead": 38,
86
+ "dead_rate": 0.0030924479166666665,
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": 5935,
93
+ "0.6": 127,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 2682,
100
+ "0.6": 63,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 6,
107
+ "0.6": 3,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.4871804714202881,
113
+ "n_dup_abs": {
114
+ "0.5": 5942,
115
+ "0.6": 195,
116
+ "0.7": 16,
117
+ "0.8": 2,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 2684,
122
+ "0.6": 97,
123
+ "0.7": 8,
124
+ "0.8": 1,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 6,
129
+ "0.6": 3,
130
+ "0.7": 2,
131
+ "0.8": 2,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.4889828860759735
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 34,
140
+ "l0_raw": 34.21214306640625,
141
+ "recon_r2": 0.898614796428707,
142
+ "l1_per_token": 186.597777168898,
143
+ "stats_source": "stats",
144
+ "n_dead": 25,
145
+ "dead_rate": 0.0020345052083333335,
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": 5966,
152
+ "0.6": 142,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 2699,
159
+ "0.6": 70,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 7,
166
+ "0.6": 3,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.48869165778160095,
172
+ "n_dup_abs": {
173
+ "0.5": 5969,
174
+ "0.6": 212,
175
+ "0.7": 14,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 2699,
181
+ "0.6": 105,
182
+ "0.7": 7,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 7,
188
+ "0.6": 3,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.49046310782432556
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.8540685176849365,
199
+ "dcs_abs": 0.8541384339332581,
200
+ "frac_sign_flipped": 0.00341796875,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 12256,
204
+ "n_alive_b": 12250,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.8559465408325195,
207
+ "0.6": 0.8553089499473572,
208
+ "0.7": 0.8564656376838684,
209
+ "0.8": 0.8564656376838684,
210
+ "0.9": 0.8564656376838684
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.7342529296875,
214
+ "0.6": 0.9919840494791666,
215
+ "0.7": 0.9971516927083334,
216
+ "0.8": 0.9971516927083334,
217
+ "0.9": 0.9971516927083334
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.6284812552621588,
221
+ "0.6": 0.8484528357245532,
222
+ "0.7": 0.8540261603629915,
223
+ "0.8": 0.8540261603629915,
224
+ "0.9": 0.8540261603629915
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.689208984375,
228
+ "0.6": 0.9131673177083334,
229
+ "0.7": 0.85888671875,
230
+ "0.8": 0.7469889322916666,
231
+ "0.9": 0.548828125
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.8558028340339661,
235
+ "0.6": 0.8556124567985535,
236
+ "0.7": 0.8565983176231384,
237
+ "0.8": 0.8565321564674377,
238
+ "0.9": 0.8565024137496948
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.7334391276041666,
242
+ "0.6": 0.9891764322916666,
243
+ "0.7": 0.99658203125,
244
+ "0.8": 0.9971110026041666,
245
+ "0.9": 0.9971516927083334
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.6276792839950455,
249
+ "0.6": 0.8463516774403008,
250
+ "0.7": 0.8536704913422,
251
+ "0.8": 0.8540576372979558,
252
+ "0.9": 0.8540628316792815
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.6878255208333334,
256
+ "0.6": 0.9100748697916666,
257
+ "0.7": 0.8588053385416666,
258
+ "0.8": 0.7470703125,
259
+ "0.9": 0.5489908854166666
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.8531638979911804,
264
+ "dcs_abs": 0.8532828688621521,
265
+ "frac_sign_flipped": 0.004069010416666667,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 12256,
269
+ "n_alive_b": 12263,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.8537924885749817,
272
+ "0.6": 0.8539842367172241,
273
+ "0.7": 0.8551971912384033,
274
+ "0.8": 0.8551971912384033,
275
+ "0.9": 0.8551971912384033
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.7342122395833334,
279
+ "0.6": 0.9921875,
280
+ "0.7": 0.9976806640625,
281
+ "0.8": 0.9976806640625,
282
+ "0.9": 0.9976806640625
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.6268648951760649,
286
+ "0.6": 0.8473124848678708,
287
+ "0.7": 0.853213701659115,
288
+ "0.8": 0.853213701659115,
289
+ "0.9": 0.853213701659115
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.6895345052083334,
293
+ "0.6": 0.914306640625,
294
+ "0.7": 0.8561197916666666,
295
+ "0.8": 0.7405598958333334,
296
+ "0.9": 0.5435384114583334
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.8535830974578857,
300
+ "0.6": 0.8543457984924316,
301
+ "0.7": 0.855301558971405,
302
+ "0.8": 0.8552937507629395,
303
+ "0.9": 0.8552937507629395
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.7334798177083334,
307
+ "0.6": 0.9893391927083334,
308
+ "0.7": 0.9971516927083334,
309
+ "0.8": 0.9976806640625,
310
+ "0.9": 0.9976806640625
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.6260859747223246,
314
+ "0.6": 0.8452377825742587,
315
+ "0.7": 0.8528653973044129,
316
+ "0.8": 0.8533100372296758,
317
+ "0.9": 0.8533100372296758
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.6883951822916666,
321
+ "0.6": 0.9117838541666666,
322
+ "0.7": 0.8553059895833334,
323
+ "0.8": 0.7406412760416666,
324
+ "0.9": 0.543701171875
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.854972779750824,
329
+ "dcs_abs": 0.8550770878791809,
330
+ "frac_sign_flipped": 0.00341796875,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 12250,
334
+ "n_alive_b": 12263,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.8539811968803406,
337
+ "0.6": 0.8557738065719604,
338
+ "0.7": 0.8573965430259705,
339
+ "0.8": 0.8573965430259705,
340
+ "0.9": 0.8573965430259705
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.7321370442708334,
344
+ "0.6": 0.9919026692708334,
345
+ "0.7": 0.9974365234375,
346
+ "0.8": 0.9974365234375,
347
+ "0.9": 0.9974365234375
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.6252312693468411,
351
+ "0.6": 0.8488443230307894,
352
+ "0.7": 0.8551986270831549,
353
+ "0.8": 0.8551986270831549,
354
+ "0.9": 0.8551986270831549
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.686767578125,
358
+ "0.6": 0.9159342447916666,
359
+ "0.7": 0.8600260416666666,
360
+ "0.8": 0.74658203125,
361
+ "0.9": 0.5498046875
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.854003369808197,
365
+ "0.6": 0.8561245799064636,
366
+ "0.7": 0.8575484752655029,
367
+ "0.8": 0.8574988842010498,
368
+ "0.9": 0.8574696779251099
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.7318115234375,
372
+ "0.6": 0.9890950520833334,
373
+ "0.7": 0.996826171875,
374
+ "0.8": 0.9973958333333334,
375
+ "0.9": 0.9974365234375
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.6249695070800954,
379
+ "0.6": 0.8467885859524056,
380
+ "0.7": 0.8548267637961544,
381
+ "0.8": 0.8552658141901096,
382
+ "0.9": 0.8552715745026944
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.6865234375,
386
+ "0.6": 0.91357421875,
387
+ "0.7": 0.8600260416666666,
388
+ "0.8": 0.7462565104166666,
389
+ "0.9": 0.5498046875
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 32,
403
+ "n_dead_std": 6.506407098647712,
404
+ "n_dense_mean": 0,
405
+ "n_dense_std": 0.0,
406
+ "l0_mean": 34,
407
+ "l0_std": 0.06916579807077423,
408
+ "recon_r2_mean": 0.8985708450103124,
409
+ "recon_r2_std": 7.01409797481315e-05,
410
+ "l1_per_token_mean": 186.09421918257712,
411
+ "l1_per_token_std": 0.5023468317282951,
412
+ "babel_mu1_signed_mean": 0.48775826891263324,
413
+ "babel_mu1_signed_std": 0.0008159380950787432,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 5925.333333333333,
416
+ "0.6": 130.66666666666666,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 32.666666666666664,
422
+ "n_dup_signed_band_std": 2.504163200219453,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 2682.6666666666665,
425
+ "0.6": 64.33333333333333,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 16.083333333333332,
431
+ "n_dup_clusters_signed_band_std": 1.282900359861721,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 6.333333333333333,
434
+ "0.6": 3.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.5,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.854068398475647,
442
+ "dcs_signed_std": 0.0009044408857138932,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.626859139928355,
445
+ "0.6": 0.8482032145410711,
446
+ "0.7": 0.8541461630350872,
447
+ "0.8": 0.8541461630350872,
448
+ "0.9": 0.8541461630350872
449
+ },
450
+ "dedup_dcs_signed_band": 0.8526604259115832,
451
+ "dedup_dcs_signed_band_std": 0.0009361319306183756,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.8545734087626139,
454
+ "0.6": 0.8550223310788473,
455
+ "0.7": 0.8563531239827474,
456
+ "0.8": 0.8563531239827474,
457
+ "0.9": 0.8563531239827474
458
+ },
459
+ "matched_cos_signed_band": 0.8560204257567724,
460
+ "matched_cos_signed_band_std": 0.0010571259962131921,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.7335340711805557,
463
+ "0.6": 0.9920247395833334,
464
+ "0.7": 0.9974229600694445,
465
+ "0.8": 0.9974229600694445,
466
+ "0.9": 0.9974229600694445
467
+ },
468
+ "unique_frac_signed_band": 0.9960734049479167,
469
+ "unique_frac_signed_band_std": 0.00022448808279124232,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.6885036892361112,
472
+ "0.6": 0.9144694010416666,
473
+ "0.7": 0.8583441840277777,
474
+ "0.8": 0.7447102864583334,
475
+ "0.9": 0.5473904079861112
476
+ },
477
+ "matched_frac_signed_band": 0.7662285698784722,
478
+ "matched_frac_signed_band_std": 0.0023179375372052525,
479
+ "babel_mu1_abs_mean": 0.48954635858535767,
480
+ "babel_mu1_abs_std": 0.0008008541951083467,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 5935.0,
483
+ "0.6": 200.0,
484
+ "0.7": 14.0,
485
+ "0.8": 0.6666666666666666,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 53.666666666666664,
489
+ "n_dup_abs_band_std": 2.6496855159307744,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 2684.6666666666665,
492
+ "0.6": 99.0,
493
+ "0.7": 7.0,
494
+ "0.8": 0.3333333333333333,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 26.583333333333332,
498
+ "n_dup_clusters_abs_band_std": 1.3768926368215255,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 6.333333333333333,
501
+ "0.6": 3.0,
502
+ "0.7": 2.0,
503
+ "0.8": 1.3333333333333333,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.8333333333333333,
507
+ "largest_dup_abs_band_std": 0.14433756729740646,
508
+ "dcs_abs_mean": 0.8541661302248637,
509
+ "dcs_abs_std": 0.0008974300996141616,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.6262449219324885,
512
+ "0.6": 0.8461260153223217,
513
+ "0.7": 0.8537875508142557,
514
+ "0.8": 0.8542111629059138,
515
+ "0.9": 0.8542148144705507
516
+ },
517
+ "dedup_dcs_abs_band": 0.8520848858782605,
518
+ "dedup_dcs_abs_band_std": 0.0009296634948064181,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.8544631004333496,
521
+ "0.6": 0.8553609450658163,
522
+ "0.7": 0.8564827839533488,
523
+ "0.8": 0.856441597143809,
524
+ "0.9": 0.856421947479248
525
+ },
526
+ "matched_cos_abs_band": 0.8561768184105555,
527
+ "matched_cos_abs_band_std": 0.0010572823834023232,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.73291015625,
530
+ "0.6": 0.9892035590277778,
531
+ "0.7": 0.9968532986111112,
532
+ "0.8": 0.9973958333333334,
533
+ "0.9": 0.9974229600694445
534
+ },
535
+ "unique_frac_abs_band": 0.9952189127604167,
536
+ "unique_frac_abs_band_std": 0.00023040266220924184,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.6875813802083334,
539
+ "0.6": 0.9118109809027777,
540
+ "0.7": 0.8580457899305555,
541
+ "0.8": 0.744656032986111,
542
+ "0.9": 0.5474989149305555
543
+ },
544
+ "matched_frac_abs_band": 0.7655029296874999,
545
+ "matched_frac_abs_band_std": 0.0023652818407118256,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 12.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 30,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 34.09826708984375,
5
+ "l0_per_token_0.001": 34.098253857421874,
6
+ "l0_per_token_0.01": 34.09807836914062,
7
+ "l0_per_token_0.05": 34.09750639648438,
8
+ "l0_per_token_0.1": 34.096228662109375,
9
+ "l0_per_token_0.5": 34.0711404296875,
10
+ "l0_per_token_1.0": 34.01213842773438,
11
+ "l1_per_token": 186.09178804915646,
12
+ "r2_score": 0.8984899544332474,
13
+ "l2_loss": 2.3806576455658055
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 12.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 30,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 34.087190966796875,
5
+ "l0_per_token_0.001": 34.0871806640625,
6
+ "l0_per_token_0.01": 34.087062158203125,
7
+ "l0_per_token_0.05": 34.086601708984375,
8
+ "l0_per_token_0.1": 34.08566435546875,
9
+ "l0_per_token_0.5": 34.06172890625,
10
+ "l0_per_token_1.0": 34.00160493164063,
11
+ "l1_per_token": 185.59309232967692,
12
+ "r2_score": 0.8986077841689825,
13
+ "l2_loss": 2.377894251463415
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 12.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 30,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 34.21214306640625,
5
+ "l0_per_token_0.001": 34.21213720703125,
6
+ "l0_per_token_0.01": 34.21210283203125,
7
+ "l0_per_token_0.05": 34.21181171875,
8
+ "l0_per_token_0.1": 34.21116733398438,
9
+ "l0_per_token_0.5": 34.19097841796875,
10
+ "l0_per_token_1.0": 34.13543442382812,
11
+ "l1_per_token": 186.597777168898,
12
+ "r2_score": 0.898614796428707,
13
+ "l2_loss": 2.3777297969049256
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_32/seed2/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 600000,
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": 68.33619833333333,
20
+ "recon_r2": 0.9322517340770707,
21
+ "l1_per_token": 255.8558036472112
22
+ },
23
+ "seed1": {
24
+ "l0": 68.25313666666666,
25
+ "recon_r2": 0.9323407862864682,
26
+ "l1_per_token": 255.34123201401795
27
+ },
28
+ "seed2": {
29
+ "l0": 68.50301666666667,
30
+ "recon_r2": 0.9322924700450567,
31
+ "l1_per_token": 256.96076816634786
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7768021305413016,
38
+ "mcc_abs": 0.7768030977677247,
39
+ "act_iou": 0.46257098043952416,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9875357167160536,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8336217505790626,
44
+ "0.7": 0.8617261810456283,
45
+ "0.8": 0.8989884177254652,
46
+ "0.9": 0.9414152728089523
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8338279584249467,
50
+ "0.7": 0.8612219114060355,
51
+ "0.8": 0.898788023707797,
52
+ "0.9": 0.9413730426214577
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.775137631695522,
58
+ "mcc_abs": 0.7751397889782531,
59
+ "act_iou": 0.46259933192133634,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.9876135734799633,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8327112898532983,
64
+ "0.7": 0.8610406097487728,
65
+ "0.8": 0.8985328214583472,
66
+ "0.9": 0.9411809905755064
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8327528887150836,
70
+ "0.7": 0.8610792045865284,
71
+ "0.8": 0.8986460702756782,
72
+ "0.9": 0.9411762675459137
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.775594728791976,
78
+ "mcc_abs": 0.7755953713512026,
79
+ "act_iou": 0.46374883008495243,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.987641363010334,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.833073454668621,
84
+ "0.7": 0.8626854635967414,
85
+ "0.8": 0.9002780058304956,
86
+ "0.9": 0.9413255147035597
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8339831061502786,
90
+ "0.7": 0.8627077226289882,
91
+ "0.8": 0.8999366437006253,
92
+ "0.9": 0.9412923738555892
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": 68.36411722222222,
106
+ "l0_std": 0.12725801297990336,
107
+ "recon_r2_mean": 0.9322949968028652,
108
+ "recon_r2_std": 4.457984273649158e-05,
109
+ "l1_per_token_mean": 256.052601275859,
110
+ "l1_per_token_std": 0.8275091039722443,
111
+ "input_r2_mean": 0.9875968844021169,
112
+ "input_r2_std": 5.4764759319457435e-05,
113
+ "mcc_signed_mean": 0.7758448303429333,
114
+ "mcc_signed_std": 0.0008599722037623513,
115
+ "mcc_abs_mean": 0.7758460860323936,
116
+ "mcc_abs_std": 0.0008595303489737128,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.833135498366994,
119
+ "0.7": 0.8618174181303808,
120
+ "0.8": 0.8992664150047694,
121
+ "0.9": 0.9413072593626728
122
+ },
123
+ "matched_mcc_signed_band": 0.8838816477162043,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8335213177634363,
126
+ "0.7": 0.8616696128738507,
127
+ "0.8": 0.8991235792280335,
128
+ "0.9": 0.9412805613409869
129
+ },
130
+ "matched_mcc_abs_band": 0.8838987678015768,
131
+ "act_iou_mean": 0.4629730474819376,
132
+ "act_iou_std": 0.0006719969769169633,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 68,
22
+ "l0_raw": 68.38207143554688,
23
+ "recon_r2": 0.9321300780911635,
24
+ "l1_per_token": 255.81543037355635,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 14,
30
+ "dense_rate": 0.0011393229166666667,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 3315,
34
+ "0.6": 2,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 1541,
41
+ "0.6": 1,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 5,
48
+ "0.6": 2,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.4383678436279297,
54
+ "n_dup_abs": {
55
+ "0.5": 3347,
56
+ "0.6": 110,
57
+ "0.7": 30,
58
+ "0.8": 2,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 1557,
63
+ "0.6": 55,
64
+ "0.7": 15,
65
+ "0.8": 1,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 5,
70
+ "0.6": 2,
71
+ "0.7": 2,
72
+ "0.8": 2,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.4412466585636139
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 68,
81
+ "l0_raw": 68.27429501953125,
82
+ "recon_r2": 0.9322215810946727,
83
+ "l1_per_token": 255.2836143987286,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 16,
89
+ "dense_rate": 0.0013020833333333333,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 3240,
93
+ "0.6": 2,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 1500,
100
+ "0.6": 1,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 7,
107
+ "0.6": 2,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.4383070468902588,
113
+ "n_dup_abs": {
114
+ "0.5": 3280,
115
+ "0.6": 98,
116
+ "0.7": 18,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 1518,
122
+ "0.6": 49,
123
+ "0.7": 9,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 7,
129
+ "0.6": 2,
130
+ "0.7": 2,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.44103512167930603
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 69,
140
+ "l0_raw": 68.547587109375,
141
+ "recon_r2": 0.9321723959031226,
142
+ "l1_per_token": 256.91514224984536,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 13,
148
+ "dense_rate": 0.0010579427083333333,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 3275,
152
+ "0.6": 2,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 1520,
159
+ "0.6": 1,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 6,
166
+ "0.6": 2,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.438535213470459,
172
+ "n_dup_abs": {
173
+ "0.5": 3318,
174
+ "0.6": 98,
175
+ "0.7": 16,
176
+ "0.8": 0,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 1541,
181
+ "0.6": 49,
182
+ "0.7": 8,
183
+ "0.8": 0,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 6,
188
+ "0.6": 2,
189
+ "0.7": 2,
190
+ "0.8": 1,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.4413144290447235
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.842194139957428,
199
+ "dcs_abs": 0.8423486351966858,
200
+ "frac_sign_flipped": 0.0035807291666666665,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 12288,
204
+ "n_alive_b": 12288,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.8376299738883972,
207
+ "0.6": 0.8421900868415833,
208
+ "0.7": 0.842194139957428,
209
+ "0.8": 0.842194139957428,
210
+ "0.9": 0.842194139957428
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.8570149739583334,
214
+ "0.6": 0.9999186197916666,
215
+ "0.7": 1.0,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.7178614302586842,
221
+ "0.6": 0.8421215492368598,
222
+ "0.7": 0.842194139957428,
223
+ "0.8": 0.842194139957428,
224
+ "0.9": 0.842194139957428
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.7935384114583334,
228
+ "0.6": 0.903076171875,
229
+ "0.7": 0.835205078125,
230
+ "0.8": 0.7178548177083334,
231
+ "0.9": 0.5221354166666666
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.8379883766174316,
235
+ "0.6": 0.8430367708206177,
236
+ "0.7": 0.8426423668861389,
237
+ "0.8": 0.8423916697502136,
238
+ "0.9": 0.8423486351966858
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.85546875,
242
+ "0.6": 0.9957682291666666,
243
+ "0.7": 0.9990234375,
244
+ "0.8": 0.9999593098958334,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.7168728690594435,
249
+ "0.6": 0.8394692324024314,
250
+ "0.7": 0.8418194739497267,
251
+ "0.8": 0.8423573927454224,
252
+ "0.9": 0.8423486351966858
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.7913411458333334,
256
+ "0.6": 0.8990885416666666,
257
+ "0.7": 0.83447265625,
258
+ "0.8": 0.71826171875,
259
+ "0.9": 0.5224609375
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.8415024280548096,
264
+ "dcs_abs": 0.8416246771812439,
265
+ "frac_sign_flipped": 0.0033365885416666665,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 12288,
269
+ "n_alive_b": 12288,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.8373281359672546,
272
+ "0.6": 0.8414965867996216,
273
+ "0.7": 0.8415024280548096,
274
+ "0.8": 0.8415024280548096,
275
+ "0.9": 0.8415024280548096
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.8564046223958334,
279
+ "0.6": 0.9999186197916666,
280
+ "0.7": 1.0,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.7170916861044437,
286
+ "0.6": 0.841428105632076,
287
+ "0.7": 0.8415024280548096,
288
+ "0.8": 0.8415024280548096,
289
+ "0.9": 0.8415024280548096
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.7937825520833334,
293
+ "0.6": 0.9007161458333334,
294
+ "0.7": 0.8323567708333334,
295
+ "0.8": 0.7145182291666666,
296
+ "0.9": 0.52197265625
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.8376181721687317,
300
+ "0.6": 0.8422266244888306,
301
+ "0.7": 0.8418663144111633,
302
+ "0.8": 0.8416663408279419,
303
+ "0.9": 0.8416246771812439
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.8548583984375,
307
+ "0.6": 0.9957682291666666,
308
+ "0.7": 0.9990641276041666,
309
+ "0.8": 0.9999593098958334,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.7160449291623081,
314
+ "0.6": 0.8386625144242619,
315
+ "0.7": 0.8410784349665239,
316
+ "0.8": 0.8416320933368601,
317
+ "0.9": 0.8416246771812439
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.79296875,
321
+ "0.6": 0.8961588541666666,
322
+ "0.7": 0.8316243489583334,
323
+ "0.8": 0.714599609375,
324
+ "0.9": 0.5220540364583334
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.8420475125312805,
329
+ "dcs_abs": 0.8422133326530457,
330
+ "frac_sign_flipped": 0.004150390625,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 12288,
334
+ "n_alive_b": 12288,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.8369420766830444,
337
+ "0.6": 0.8420504331588745,
338
+ "0.7": 0.8420475125312805,
339
+ "0.8": 0.8420475125312805,
340
+ "0.9": 0.8420475125312805
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.8577880859375,
344
+ "0.6": 0.9999186197916666,
345
+ "0.7": 1.0,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.717918941998505,
351
+ "0.6": 0.8419819069191968,
352
+ "0.7": 0.8420475125312805,
353
+ "0.8": 0.8420475125312805,
354
+ "0.9": 0.8420475125312805
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.7963053385416666,
358
+ "0.6": 0.9016927083333334,
359
+ "0.7": 0.8310546875,
360
+ "0.8": 0.7145182291666666,
361
+ "0.9": 0.526611328125
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.8375152945518494,
365
+ "0.6": 0.8427015542984009,
366
+ "0.7": 0.8422761559486389,
367
+ "0.8": 0.8422133326530457,
368
+ "0.9": 0.8422133326530457
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.8559977213541666,
372
+ "0.6": 0.9960123697916666,
373
+ "0.7": 0.9993082682291666,
374
+ "0.8": 1.0,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.7169111837356468,
379
+ "0.6": 0.839341172123871,
380
+ "0.7": 0.8416935267717538,
381
+ "0.8": 0.8422133326530457,
382
+ "0.9": 0.8422133326530457
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.794921875,
386
+ "0.6": 0.8975423177083334,
387
+ "0.7": 0.8302408854166666,
388
+ "0.8": 0.7154134114583334,
389
+ "0.9": 0.5266927083333334
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 14,
405
+ "n_dense_std": 1.5275252316519468,
406
+ "l0_mean": 68,
407
+ "l0_std": 0.13766530314758454,
408
+ "recon_r2_mean": 0.932174685029653,
409
+ "recon_r2_std": 3.464368346113687e-05,
410
+ "l1_per_token_mean": 256.00472900737674,
411
+ "l1_per_token_std": 0.8320763648757366,
412
+ "babel_mu1_signed_mean": 0.4384033679962158,
413
+ "babel_mu1_signed_std": 0.00011815871798838899,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 3276.6666666666665,
416
+ "0.6": 2.0,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 0.5,
422
+ "n_dup_signed_band_std": 0.0,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 1520.3333333333333,
425
+ "0.6": 1.0,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 0.25,
431
+ "n_dup_clusters_signed_band_std": 0.0,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 6.0,
434
+ "0.6": 2.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.25,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.841914693514506,
442
+ "dcs_signed_std": 0.00036448183420947374,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.7176240194538778,
445
+ "0.6": 0.8418438539293774,
446
+ "0.7": 0.841914693514506,
447
+ "0.8": 0.841914693514506,
448
+ "0.9": 0.841914693514506
449
+ },
450
+ "dedup_dcs_signed_band": 0.8418969836182238,
451
+ "dedup_dcs_signed_band_std": 0.00036504624021942595,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.8373000621795654,
454
+ "0.6": 0.8419123689333597,
455
+ "0.7": 0.841914693514506,
456
+ "0.8": 0.841914693514506,
457
+ "0.9": 0.841914693514506
458
+ },
459
+ "matched_cos_signed_band": 0.8419141123692194,
460
+ "matched_cos_signed_band_std": 0.00036505370224472965,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.8570692274305557,
463
+ "0.6": 0.9999186197916666,
464
+ "0.7": 1.0,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 0.9999796549479166,
469
+ "unique_frac_signed_band_std": 0.0,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.7945421006944445,
472
+ "0.6": 0.901828342013889,
473
+ "0.7": 0.8328721788194445,
474
+ "0.8": 0.7156304253472222,
475
+ "0.9": 0.5235731336805555
476
+ },
477
+ "matched_frac_signed_band": 0.7434760199652778,
478
+ "matched_frac_signed_band_std": 0.0010884761313991254,
479
+ "babel_mu1_abs_mean": 0.4411987364292145,
480
+ "babel_mu1_abs_std": 0.00014568990808618186,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 3315.0,
483
+ "0.6": 102.0,
484
+ "0.7": 21.333333333333332,
485
+ "0.8": 0.6666666666666666,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 31.0,
489
+ "n_dup_abs_band_std": 3.905124837953327,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 1538.6666666666667,
492
+ "0.6": 51.0,
493
+ "0.7": 10.666666666666666,
494
+ "0.8": 0.3333333333333333,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 15.5,
498
+ "n_dup_clusters_abs_band_std": 1.9525624189766635,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 6.0,
501
+ "0.6": 2.0,
502
+ "0.7": 2.0,
503
+ "0.8": 1.3333333333333333,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.5833333333333333,
507
+ "largest_dup_abs_band_std": 0.14433756729740646,
508
+ "dcs_abs_mean": 0.8420622150103251,
509
+ "dcs_abs_std": 0.00038491065000425916,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.7166096606524661,
512
+ "0.6": 0.8391576396501881,
513
+ "0.7": 0.8415304785626682,
514
+ "0.8": 0.8420676062451093,
515
+ "0.9": 0.8420622150103251
516
+ },
517
+ "dedup_dcs_abs_band": 0.8412044848670727,
518
+ "dedup_dcs_abs_band_std": 0.0003996889687014163,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.8377072811126709,
521
+ "0.6": 0.8426549832026163,
522
+ "0.7": 0.8422616124153137,
523
+ "0.8": 0.8420904477437338,
524
+ "0.9": 0.8420622150103251
525
+ },
526
+ "matched_cos_abs_band": 0.8422673145929972,
527
+ "matched_cos_abs_band_std": 0.00038631035693890634,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.8554416232638888,
530
+ "0.6": 0.995849609375,
531
+ "0.7": 0.9991319444444443,
532
+ "0.8": 0.999972873263889,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.9987386067708333,
536
+ "unique_frac_abs_band_std": 7.944996822004592e-05,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.7930772569444445,
539
+ "0.6": 0.8975965711805555,
540
+ "0.7": 0.8321126302083334,
541
+ "0.8": 0.7160915798611112,
542
+ "0.9": 0.5237358940972223
543
+ },
544
+ "matched_frac_abs_band": 0.7423841688368056,
545
+ "matched_frac_abs_band_std": 0.0012332413623357275,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 12288
549
+ }
550
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 6.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 18,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 68.38207143554688,
5
+ "l0_per_token_0.001": 68.3819552734375,
6
+ "l0_per_token_0.01": 68.38150571289063,
7
+ "l0_per_token_0.05": 68.37984819335938,
8
+ "l0_per_token_0.1": 68.37665571289062,
9
+ "l0_per_token_0.5": 68.29212431640624,
10
+ "l0_per_token_1.0": 63.50503798828125,
11
+ "l1_per_token": 255.81543037355635,
12
+ "r2_score": 0.9321300780911635,
13
+ "l2_loss": 1.5917148652048885
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed0/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 6.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 18,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 68.27429501953125,
5
+ "l0_per_token_0.001": 68.27426918945312,
6
+ "l0_per_token_0.01": 68.27396088867188,
7
+ "l0_per_token_0.05": 68.27259619140625,
8
+ "l0_per_token_0.1": 68.269659375,
9
+ "l0_per_token_0.5": 68.191526171875,
10
+ "l0_per_token_1.0": 63.4634873046875,
11
+ "l1_per_token": 255.2836143987286,
12
+ "r2_score": 0.9322215810946727,
13
+ "l2_loss": 1.5895688970528683
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed1/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 6.5,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 18,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 68.547587109375,
5
+ "l0_per_token_0.001": 68.5475615234375,
6
+ "l0_per_token_0.01": 68.54738310546875,
7
+ "l0_per_token_0.05": 68.54639868164062,
8
+ "l0_per_token_0.1": 68.54444653320313,
9
+ "l0_per_token_0.5": 68.47152568359375,
10
+ "l0_per_token_1.0": 63.59962797851563,
11
+ "l1_per_token": 256.91514224984536,
12
+ "r2_score": 0.9321723959031226,
13
+ "l2_loss": 1.5907224095122363
14
+ }
lm/gpt2-small_layer8/consdl/x16/l0_64/seed2/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 12288,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 600000,
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": 127.844745,
20
+ "recon_r2": 0.9415145041736372,
21
+ "l1_per_token": 393.4831958380243
22
+ },
23
+ "seed1": {
24
+ "l0": 127.82662833333333,
25
+ "recon_r2": 0.9413935058689876,
26
+ "l1_per_token": 392.6340377780998
27
+ },
28
+ "seed2": {
29
+ "l0": 127.94816,
30
+ "recon_r2": 0.9412829085142881,
31
+ "l1_per_token": 392.29133534548754
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.8384991348521575,
38
+ "mcc_abs": 0.8385107009141827,
39
+ "act_iou": 0.5043471564033918,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9880527451632783,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8957322305283675,
44
+ "0.7": 0.905955677007102,
45
+ "0.8": 0.9195577963035144,
46
+ "0.9": 0.9397493154646301
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8954973345640642,
50
+ "0.7": 0.904681714997588,
51
+ "0.8": 0.9186516445179674,
52
+ "0.9": 0.9396600462816262
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.8343279700795557,
58
+ "mcc_abs": 0.834332741334677,
59
+ "act_iou": 0.49883932105666623,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9878843753705879,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8940291203862878,
64
+ "0.7": 0.9052258916690518,
65
+ "0.8": 0.9194332997748622,
66
+ "0.9": 0.9378714317629683
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.892954798378118,
70
+ "0.7": 0.902677393241714,
71
+ "0.8": 0.9187113669722669,
72
+ "0.9": 0.937871431762968
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8358634571464081,
78
+ "mcc_abs": 0.8358725738918551,
79
+ "act_iou": 0.4990281240433236,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9878902078566736,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8962142435565161,
84
+ "0.7": 0.9058219536751124,
85
+ "0.8": 0.9194503320038773,
86
+ "0.9": 0.9381452531661661
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8950302618959097,
90
+ "0.7": 0.9037448659876602,
91
+ "0.8": 0.9182291337122168,
92
+ "0.9": 0.9375700971776684
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": 127.8731777777778,
106
+ "l0_std": 0.06556526244754074,
107
+ "recon_r2_mean": 0.9413969728523043,
108
+ "recon_r2_std": 0.00011583674864796442,
109
+ "l1_per_token_mean": 392.8028563205373,
110
+ "l1_per_token_std": 0.6136021786308381,
111
+ "input_r2_mean": 0.9879424427968466,
112
+ "input_r2_std": 9.556915560898149e-05,
113
+ "mcc_signed_mean": 0.8362301873593738,
114
+ "mcc_signed_std": 0.0021096260751315627,
115
+ "mcc_abs_mean": 0.8362386720469049,
116
+ "mcc_abs_std": 0.002112902614022678,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8953251981570571,
119
+ "0.7": 0.9056678407837554,
120
+ "0.8": 0.919480476027418,
121
+ "0.9": 0.9385886667979215
122
+ },
123
+ "matched_mcc_signed_band": 0.914765545441538,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8944941316126972,
126
+ "0.7": 0.9037013247423208,
127
+ "0.8": 0.9185307150674836,
128
+ "0.9": 0.9383671917407543
129
+ },
130
+ "matched_mcc_abs_band": 0.913773340790814,
131
+ "act_iou_mean": 0.5007382005011273,
132
+ "act_iou_std": 0.0031268728260532867,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,550 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 3072,
20
+ "d_in": 768,
21
+ "l0": 126,
22
+ "l0_raw": 125.704651171875,
23
+ "recon_r2": 0.9411897999822579,
24
+ "l1_per_token": 391.79768565147,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 51,
30
+ "dense_rate": 0.0166015625,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 229,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 109,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 3,
48
+ "0.6": 1,
49
+ "0.7": 1,
50
+ "0.8": 1,
51
+ "0.9": 1
52
+ },
53
+ "babel_mu1_signed": 0.33925139904022217,
54
+ "n_dup_abs": {
55
+ "0.5": 305,
56
+ "0.6": 88,
57
+ "0.7": 54,
58
+ "0.8": 10,
59
+ "0.9": 0
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 147,
63
+ "0.6": 44,
64
+ "0.7": 27,
65
+ "0.8": 5,
66
+ "0.9": 0
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 3,
70
+ "0.6": 2,
71
+ "0.7": 2,
72
+ "0.8": 2,
73
+ "0.9": 1
74
+ },
75
+ "babel_mu1_abs": 0.35641753673553467
76
+ },
77
+ "seed1": {
78
+ "d_sae": 3072,
79
+ "d_in": 768,
80
+ "l0": 126,
81
+ "l0_raw": 125.63611196289062,
82
+ "recon_r2": 0.9410739181284478,
83
+ "l1_per_token": 390.88694813323394,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 53,
89
+ "dense_rate": 0.017252604166666668,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 222,
93
+ "0.6": 0,
94
+ "0.7": 0,
95
+ "0.8": 0,
96
+ "0.9": 0
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 105,
100
+ "0.6": 0,
101
+ "0.7": 0,
102
+ "0.8": 0,
103
+ "0.9": 0
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 4,
107
+ "0.6": 1,
108
+ "0.7": 1,
109
+ "0.8": 1,
110
+ "0.9": 1
111
+ },
112
+ "babel_mu1_signed": 0.3414287567138672,
113
+ "n_dup_abs": {
114
+ "0.5": 296,
115
+ "0.6": 80,
116
+ "0.7": 46,
117
+ "0.8": 10,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 143,
122
+ "0.6": 40,
123
+ "0.7": 23,
124
+ "0.8": 5,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 4,
129
+ "0.6": 2,
130
+ "0.7": 2,
131
+ "0.8": 2,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_abs": 0.35803985595703125
135
+ },
136
+ "seed2": {
137
+ "d_sae": 3072,
138
+ "d_in": 768,
139
+ "l0": 126,
140
+ "l0_raw": 125.77726040039063,
141
+ "recon_r2": 0.9409605591020928,
142
+ "l1_per_token": 390.5783639531069,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 55,
148
+ "dense_rate": 0.017903645833333332,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 212,
152
+ "0.6": 0,
153
+ "0.7": 0,
154
+ "0.8": 0,
155
+ "0.9": 0
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 101,
159
+ "0.6": 0,
160
+ "0.7": 0,
161
+ "0.8": 0,
162
+ "0.9": 0
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 4,
166
+ "0.6": 1,
167
+ "0.7": 1,
168
+ "0.8": 1,
169
+ "0.9": 1
170
+ },
171
+ "babel_mu1_signed": 0.3400411605834961,
172
+ "n_dup_abs": {
173
+ "0.5": 272,
174
+ "0.6": 78,
175
+ "0.7": 44,
176
+ "0.8": 6,
177
+ "0.9": 0
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 131,
181
+ "0.6": 39,
182
+ "0.7": 22,
183
+ "0.8": 3,
184
+ "0.9": 0
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 4,
188
+ "0.6": 2,
189
+ "0.7": 2,
190
+ "0.8": 2,
191
+ "0.9": 1
192
+ },
193
+ "babel_mu1_abs": 0.3550417423248291
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.8878882527351379,
199
+ "dcs_abs": 0.8885080218315125,
200
+ "frac_sign_flipped": 0.012044270833333334,
201
+ "alive_source": "stats",
202
+ "frac_denominator": "d_sae",
203
+ "n_alive_a": 3072,
204
+ "n_alive_b": 3072,
205
+ "matched_cos_signed": {
206
+ "0.5": 0.8854705691337585,
207
+ "0.6": 0.8878882527351379,
208
+ "0.7": 0.8878882527351379,
209
+ "0.8": 0.8878882527351379,
210
+ "0.9": 0.8878882527351379
211
+ },
212
+ "unique_frac_signed": {
213
+ "0.5": 0.96142578125,
214
+ "0.6": 1.0,
215
+ "0.7": 1.0,
216
+ "0.8": 1.0,
217
+ "0.9": 1.0
218
+ },
219
+ "dedup_dcs_signed": {
220
+ "0.5": 0.851314233703306,
221
+ "0.6": 0.8878882527351379,
222
+ "0.7": 0.8878882527351379,
223
+ "0.8": 0.8878882527351379,
224
+ "0.9": 0.8878882527351379
225
+ },
226
+ "matched_frac_signed": {
227
+ "0.5": 0.890625,
228
+ "0.6": 0.9153645833333334,
229
+ "0.7": 0.8896484375,
230
+ "0.8": 0.8453776041666666,
231
+ "0.9": 0.7454427083333334
232
+ },
233
+ "matched_cos_abs": {
234
+ "0.5": 0.8882692456245422,
235
+ "0.6": 0.8909647464752197,
236
+ "0.7": 0.8901378512382507,
237
+ "0.8": 0.8887239098548889,
238
+ "0.9": 0.8885080218315125
239
+ },
240
+ "unique_frac_abs": {
241
+ "0.5": 0.9493815104166666,
242
+ "0.6": 0.986328125,
243
+ "0.7": 0.9918619791666666,
244
+ "0.8": 0.9983723958333334,
245
+ "0.9": 1.0
246
+ },
247
+ "dedup_dcs_abs": {
248
+ "0.5": 0.8433063980677009,
249
+ "0.6": 0.8787835878320038,
250
+ "0.7": 0.8828938908603352,
251
+ "0.8": 0.8872774191161928,
252
+ "0.9": 0.8885080218315125
253
+ },
254
+ "matched_frac_abs": {
255
+ "0.5": 0.8802083333333334,
256
+ "0.6": 0.904296875,
257
+ "0.7": 0.8841145833333334,
258
+ "0.8": 0.845703125,
259
+ "0.9": 0.74609375
260
+ }
261
+ },
262
+ "seed0\u2194seed2": {
263
+ "dcs_signed": 0.8845382332801819,
264
+ "dcs_abs": 0.8851227760314941,
265
+ "frac_sign_flipped": 0.014322916666666666,
266
+ "alive_source": "stats",
267
+ "frac_denominator": "d_sae",
268
+ "n_alive_a": 3072,
269
+ "n_alive_b": 3072,
270
+ "matched_cos_signed": {
271
+ "0.5": 0.8836352229118347,
272
+ "0.6": 0.8845381736755371,
273
+ "0.7": 0.8845381736755371,
274
+ "0.8": 0.8845381736755371,
275
+ "0.9": 0.8845381736755371
276
+ },
277
+ "unique_frac_signed": {
278
+ "0.5": 0.96240234375,
279
+ "0.6": 1.0,
280
+ "0.7": 1.0,
281
+ "0.8": 1.0,
282
+ "0.9": 1.0
283
+ },
284
+ "dedup_dcs_signed": {
285
+ "0.5": 0.8504126095504034,
286
+ "0.6": 0.8845381736755371,
287
+ "0.7": 0.8845381736755371,
288
+ "0.8": 0.8845381736755371,
289
+ "0.9": 0.8845381736755371
290
+ },
291
+ "matched_frac_signed": {
292
+ "0.5": 0.8922526041666666,
293
+ "0.6": 0.908203125,
294
+ "0.7": 0.8824869791666666,
295
+ "0.8": 0.8382161458333334,
296
+ "0.9": 0.74609375
297
+ },
298
+ "matched_cos_abs": {
299
+ "0.5": 0.8872216939926147,
300
+ "0.6": 0.8873640298843384,
301
+ "0.7": 0.8865522742271423,
302
+ "0.8": 0.8858667016029358,
303
+ "0.9": 0.8851227760314941
304
+ },
305
+ "unique_frac_abs": {
306
+ "0.5": 0.9513346354166666,
307
+ "0.6": 0.9864908854166666,
308
+ "0.7": 0.9920247395833334,
309
+ "0.8": 0.9986979166666666,
310
+ "0.9": 1.0
311
+ },
312
+ "dedup_dcs_abs": {
313
+ "0.5": 0.8440447267882215,
314
+ "0.6": 0.8753765275275024,
315
+ "0.7": 0.8794817889671929,
316
+ "0.8": 0.8847132293352236,
317
+ "0.9": 0.8851227760314941
318
+ },
319
+ "matched_frac_abs": {
320
+ "0.5": 0.8831380208333334,
321
+ "0.6": 0.8955078125,
322
+ "0.7": 0.8779296875,
323
+ "0.8": 0.8382161458333334,
324
+ "0.9": 0.74609375
325
+ }
326
+ },
327
+ "seed1\u2194seed2": {
328
+ "dcs_signed": 0.8841626644134521,
329
+ "dcs_abs": 0.884857714176178,
330
+ "frac_sign_flipped": 0.012369791666666666,
331
+ "alive_source": "stats",
332
+ "frac_denominator": "d_sae",
333
+ "n_alive_a": 3072,
334
+ "n_alive_b": 3072,
335
+ "matched_cos_signed": {
336
+ "0.5": 0.8812102675437927,
337
+ "0.6": 0.8841626644134521,
338
+ "0.7": 0.8841626644134521,
339
+ "0.8": 0.8841626644134521,
340
+ "0.9": 0.8841626644134521
341
+ },
342
+ "unique_frac_signed": {
343
+ "0.5": 0.962890625,
344
+ "0.6": 1.0,
345
+ "0.7": 1.0,
346
+ "0.8": 1.0,
347
+ "0.9": 1.0
348
+ },
349
+ "dedup_dcs_signed": {
350
+ "0.5": 0.8485091052716598,
351
+ "0.6": 0.8841626644134521,
352
+ "0.7": 0.8841626644134521,
353
+ "0.8": 0.8841626644134521,
354
+ "0.9": 0.8841626644134521
355
+ },
356
+ "matched_frac_signed": {
357
+ "0.5": 0.8870442708333334,
358
+ "0.6": 0.9065755208333334,
359
+ "0.7": 0.8834635416666666,
360
+ "0.8": 0.8408203125,
361
+ "0.9": 0.7425130208333334
362
+ },
363
+ "matched_cos_abs": {
364
+ "0.5": 0.8841249346733093,
365
+ "0.6": 0.8862740993499756,
366
+ "0.7": 0.8853506445884705,
367
+ "0.8": 0.8853362202644348,
368
+ "0.9": 0.884857714176178
369
+ },
370
+ "unique_frac_abs": {
371
+ "0.5": 0.9521484375,
372
+ "0.6": 0.9871419270833334,
373
+ "0.7": 0.99267578125,
374
+ "0.8": 0.9986979166666666,
375
+ "0.9": 1.0
376
+ },
377
+ "dedup_dcs_abs": {
378
+ "0.5": 0.841818175103981,
379
+ "0.6": 0.8748783223563805,
380
+ "0.7": 0.878866142797051,
381
+ "0.8": 0.8841834387276322,
382
+ "0.9": 0.884857714176178
383
+ },
384
+ "matched_frac_abs": {
385
+ "0.5": 0.8772786458333334,
386
+ "0.6": 0.896484375,
387
+ "0.7": 0.8792317708333334,
388
+ "0.8": 0.841796875,
389
+ "0.9": 0.7425130208333334
390
+ }
391
+ }
392
+ },
393
+ "summary": {
394
+ "n_seeds": 3,
395
+ "n_pairs": 3,
396
+ "report_band": [
397
+ 0.6,
398
+ 0.7,
399
+ 0.8,
400
+ 0.9
401
+ ],
402
+ "n_dead_mean": 0,
403
+ "n_dead_std": 0.0,
404
+ "n_dense_mean": 53,
405
+ "n_dense_std": 2.0,
406
+ "l0_mean": 126,
407
+ "l0_std": 0.0705839980006476,
408
+ "recon_r2_mean": 0.9410747590709327,
409
+ "recon_r2_std": 0.00011462275373120408,
410
+ "l1_per_token_mean": 391.0876659126036,
411
+ "l1_per_token_std": 0.6339574680063991,
412
+ "babel_mu1_signed_mean": 0.34024043877919513,
413
+ "babel_mu1_signed_std": 0.0011022728605957857,
414
+ "n_dup_signed_per_threshold": {
415
+ "0.5": 221.0,
416
+ "0.6": 0.0,
417
+ "0.7": 0.0,
418
+ "0.8": 0.0,
419
+ "0.9": 0.0
420
+ },
421
+ "n_dup_signed_band": 0.0,
422
+ "n_dup_signed_band_std": 0.0,
423
+ "n_dup_clusters_signed_per_threshold": {
424
+ "0.5": 105.0,
425
+ "0.6": 0.0,
426
+ "0.7": 0.0,
427
+ "0.8": 0.0,
428
+ "0.9": 0.0
429
+ },
430
+ "n_dup_clusters_signed_band": 0.0,
431
+ "n_dup_clusters_signed_band_std": 0.0,
432
+ "largest_dup_signed_per_threshold": {
433
+ "0.5": 3.6666666666666665,
434
+ "0.6": 1.0,
435
+ "0.7": 1.0,
436
+ "0.8": 1.0,
437
+ "0.9": 1.0
438
+ },
439
+ "largest_dup_signed_band": 1.0,
440
+ "largest_dup_signed_band_std": 0.0,
441
+ "dcs_signed_mean": 0.8855297168095907,
442
+ "dcs_signed_std": 0.002051165955618181,
443
+ "dedup_dcs_signed_per_threshold": {
444
+ "0.5": 0.8500786495084564,
445
+ "0.6": 0.8855296969413757,
446
+ "0.7": 0.8855296969413757,
447
+ "0.8": 0.8855296969413757,
448
+ "0.9": 0.8855296969413757
449
+ },
450
+ "dedup_dcs_signed_band": 0.8855296969413757,
451
+ "dedup_dcs_signed_band_std": 0.002051180361571077,
452
+ "matched_cos_signed_per_threshold": {
453
+ "0.5": 0.8834386865297953,
454
+ "0.6": 0.8855296969413757,
455
+ "0.7": 0.8855296969413757,
456
+ "0.8": 0.8855296969413757,
457
+ "0.9": 0.8855296969413757
458
+ },
459
+ "matched_cos_signed_band": 0.8855296969413757,
460
+ "matched_cos_signed_band_std": 0.002051180361571077,
461
+ "unique_frac_signed_per_threshold": {
462
+ "0.5": 0.9622395833333334,
463
+ "0.6": 1.0,
464
+ "0.7": 1.0,
465
+ "0.8": 1.0,
466
+ "0.9": 1.0
467
+ },
468
+ "unique_frac_signed_band": 1.0,
469
+ "unique_frac_signed_band_std": 0.0,
470
+ "matched_frac_signed_per_threshold": {
471
+ "0.5": 0.8899739583333334,
472
+ "0.6": 0.9100477430555557,
473
+ "0.7": 0.8851996527777777,
474
+ "0.8": 0.8414713541666666,
475
+ "0.9": 0.7446831597222223
476
+ },
477
+ "matched_frac_signed_band": 0.8453504774305556,
478
+ "matched_frac_signed_band_std": 0.003131111667736198,
479
+ "babel_mu1_abs_mean": 0.356499711672465,
480
+ "babel_mu1_abs_std": 0.0015007451076361096,
481
+ "n_dup_abs_per_threshold": {
482
+ "0.5": 291.0,
483
+ "0.6": 82.0,
484
+ "0.7": 48.0,
485
+ "0.8": 8.666666666666666,
486
+ "0.9": 0.0
487
+ },
488
+ "n_dup_abs_band": 34.666666666666664,
489
+ "n_dup_abs_band_std": 3.055050463303893,
490
+ "n_dup_clusters_abs_per_threshold": {
491
+ "0.5": 140.33333333333334,
492
+ "0.6": 41.0,
493
+ "0.7": 24.0,
494
+ "0.8": 4.333333333333333,
495
+ "0.9": 0.0
496
+ },
497
+ "n_dup_clusters_abs_band": 17.333333333333332,
498
+ "n_dup_clusters_abs_band_std": 1.5275252316519465,
499
+ "largest_dup_abs_per_threshold": {
500
+ "0.5": 3.6666666666666665,
501
+ "0.6": 2.0,
502
+ "0.7": 2.0,
503
+ "0.8": 2.0,
504
+ "0.9": 1.0
505
+ },
506
+ "largest_dup_abs_band": 1.75,
507
+ "largest_dup_abs_band_std": 0.0,
508
+ "dcs_abs_mean": 0.8861628373463949,
509
+ "dcs_abs_std": 0.0020353088583094994,
510
+ "dedup_dcs_abs_per_threshold": {
511
+ "0.5": 0.8430564333199678,
512
+ "0.6": 0.8763461459052956,
513
+ "0.7": 0.8804139408748597,
514
+ "0.8": 0.8853913623930162,
515
+ "0.9": 0.8861628373463949
516
+ },
517
+ "dedup_dcs_abs_band": 0.8820785716298917,
518
+ "dedup_dcs_abs_band_std": 0.001995054879145602,
519
+ "matched_cos_abs_per_threshold": {
520
+ "0.5": 0.8865386247634888,
521
+ "0.6": 0.8882009585698446,
522
+ "0.7": 0.8873469233512878,
523
+ "0.8": 0.8866422772407532,
524
+ "0.9": 0.8861628373463949
525
+ },
526
+ "matched_cos_abs_band": 0.8870882491270702,
527
+ "matched_cos_abs_band_std": 0.0021952477216257327,
528
+ "unique_frac_abs_per_threshold": {
529
+ "0.5": 0.950954861111111,
530
+ "0.6": 0.9866536458333334,
531
+ "0.7": 0.9921875,
532
+ "0.8": 0.9985894097222222,
533
+ "0.9": 1.0
534
+ },
535
+ "unique_frac_abs_band": 0.994357638888889,
536
+ "unique_frac_abs_band_std": 0.0002486206431725214,
537
+ "matched_frac_abs_per_threshold": {
538
+ "0.5": 0.8802083333333334,
539
+ "0.6": 0.8987630208333334,
540
+ "0.7": 0.8804253472222223,
541
+ "0.8": 0.8419053819444445,
542
+ "0.9": 0.7449001736111112
543
+ },
544
+ "matched_frac_abs_band": 0.8414984809027779,
545
+ "matched_frac_abs_band_std": 0.003090662747715338,
546
+ "alive_source": "stats",
547
+ "frac_denominator": "d_sae",
548
+ "d_sae": 3072
549
+ }
550
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.95,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 125.704651171875,
5
+ "l0_per_token_0.001": 125.7046107421875,
6
+ "l0_per_token_0.01": 125.7044861328125,
7
+ "l0_per_token_0.05": 125.70372827148438,
8
+ "l0_per_token_0.1": 125.70257119140625,
9
+ "l0_per_token_0.5": 125.60522626953124,
10
+ "l0_per_token_1.0": 94.27361411132813,
11
+ "l1_per_token": 391.79768565147,
12
+ "r2_score": 0.9411897999822579,
13
+ "l2_loss": 1.3792423353551146
14
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 3072,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.95,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 125.63611196289062,
5
+ "l0_per_token_0.001": 125.63610112304687,
6
+ "l0_per_token_0.01": 125.63594052734375,
7
+ "l0_per_token_0.05": 125.635241796875,
8
+ "l0_per_token_0.1": 125.63404838867187,
9
+ "l0_per_token_0.5": 125.52499116210937,
10
+ "l0_per_token_1.0": 93.74265239257812,
11
+ "l1_per_token": 390.88694813323394,
12
+ "r2_score": 0.9410739181284478,
13
+ "l2_loss": 1.3819600468851936
14
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 3072,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 2.95,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 5,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 125.77726040039063,
5
+ "l0_per_token_0.001": 125.77724150390625,
6
+ "l0_per_token_0.01": 125.77699580078125,
7
+ "l0_per_token_0.05": 125.77617680664062,
8
+ "l0_per_token_0.1": 125.774969140625,
9
+ "l0_per_token_0.5": 125.6611578125,
10
+ "l0_per_token_1.0": 93.548204296875,
11
+ "l1_per_token": 390.5783639531069,
12
+ "r2_score": 0.9409605591020928,
13
+ "l2_loss": 1.384618591970846
14
+ }
lm/gpt2-small_layer8/consdl/x4/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 3072,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x4/l0_32/code_metrics_step_final.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 1,
3
+ "step": "final",
4
+ "n_tokens": 600000,
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": 33.676136666666665,
20
+ "recon_r2": 0.8975673795678348,
21
+ "l1_per_token": 253.82649805907457
22
+ },
23
+ "seed1": {
24
+ "l0": 33.846435,
25
+ "recon_r2": 0.8977308049876321,
26
+ "l1_per_token": 254.91719231354833
27
+ },
28
+ "seed2": {
29
+ "l0": 33.651356666666665,
30
+ "recon_r2": 0.8976816471087172,
31
+ "l1_per_token": 255.21824300224597
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.8431015566494878,
38
+ "mcc_abs": 0.8431068267004274,
39
+ "act_iou": 0.5771575229163576,
40
+ "act_iou_n_pairs": 3072,
41
+ "input_r2": 0.9810264992004553,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8933773801726083,
44
+ "0.7": 0.9059894971164321,
45
+ "0.8": 0.9244591009457132,
46
+ "0.9": 0.9450882109904006
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8941537238424455,
50
+ "0.7": 0.9051395484599205,
51
+ "0.8": 0.9227457160326656,
52
+ "0.9": 0.9450882109904014
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.8272049425533208,
58
+ "mcc_abs": 0.8272087362405648,
59
+ "act_iou": 0.5592969803272019,
60
+ "act_iou_n_pairs": 3072,
61
+ "input_r2": 0.9800477389075511,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8880478491749985,
64
+ "0.7": 0.9024123006077128,
65
+ "0.8": 0.9184008541940856,
66
+ "0.9": 0.9426502364837515
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8882316570609861,
70
+ "0.7": 0.9017869623984147,
71
+ "0.8": 0.9181568606864884,
72
+ "0.9": 0.9426502364837515
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.8336904002024171,
78
+ "mcc_abs": 0.8337068280859405,
79
+ "act_iou": 0.5680103794159135,
80
+ "act_iou_n_pairs": 3072,
81
+ "input_r2": 0.9805106627129005,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8906078080022163,
84
+ "0.7": 0.9041514540728112,
85
+ "0.8": 0.9187560352020746,
86
+ "0.9": 0.9430678425864627
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8897145007397163,
90
+ "0.7": 0.9036372232099904,
91
+ "0.8": 0.917934181590078,
92
+ "0.9": 0.9430055361585956
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": 33.724642777777774,
106
+ "l0_std": 0.10620038203966387,
107
+ "recon_r2_mean": 0.8976599438880614,
108
+ "recon_r2_std": 8.384652228391995e-05,
109
+ "l1_per_token_mean": 254.65397779162296,
110
+ "l1_per_token_std": 0.7322567241924753,
111
+ "input_r2_mean": 0.9805283002736357,
112
+ "input_r2_std": 0.0004896184641156311,
113
+ "mcc_signed_mean": 0.8346656331350752,
114
+ "mcc_signed_std": 0.007993052882680055,
115
+ "mcc_abs_mean": 0.8346741303423109,
116
+ "mcc_abs_std": 0.007993064200222446,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8906776791166077,
119
+ "0.7": 0.904184417265652,
120
+ "0.8": 0.9205386634472911,
121
+ "0.9": 0.9436020966868716
122
+ },
123
+ "matched_mcc_signed_band": 0.9147507141291056,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.890699960547716,
126
+ "0.7": 0.903521244689442,
127
+ "0.8": 0.919612252769744,
128
+ "0.9": 0.9435813278775829
129
+ },
130
+ "matched_mcc_abs_band": 0.9143536964711212,
131
+ "act_iou_mean": 0.568154960886491,
132
+ "act_iou_std": 0.008931149043991313,
133
+ "act_iou_n_pairs": 3072.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/consdl/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": 34,
22
+ "l0_raw": 34.167963525390626,
23
+ "recon_r2": 0.8977384601466455,
24
+ "l1_per_token": 255.04705028489641,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 20,
30
+ "dense_rate": 0.006510416666666667,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 313,
34
+ "0.6": 0,
35
+ "0.7": 0,
36
+ "0.8": 0,
37
+ "0.9": 0
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 149,
41
+ "0.6": 0,
42
+ "0.7": 0,
43
+ "0.8": 0,
44
+ "0.9": 0
45
+ },
46
+ "n_dup_removed_signed": {
47
+ "0.5": 164,
48
+ "0.6": 0,
49
+ "0.7": 0,
50
+ "0.8": 0,
51
+ "0.9": 0
52
+ },
53
+ "largest_dup_signed": {
54
+ "0.5": 4,
55
+ "0.6": 1,
56
+ "0.7": 1,
57
+ "0.8": 1,
58
+ "0.9": 1
59
+ },
60
+ "babel_mu1_signed": 0.37319397926330566,
61
+ "n_dup_abs": {
62
+ "0.5": 328,
63
+ "0.6": 40,
64
+ "0.7": 26,
65
+ "0.8": 8,
66
+ "0.9": 0
67
+ },
68
+ "n_dup_clusters_abs": {
69
+ "0.5": 155,
70
+ "0.6": 20,
71
+ "0.7": 13,
72
+ "0.8": 4,
73
+ "0.9": 0
74
+ },
75
+ "n_dup_removed_abs": {
76
+ "0.5": 173,
77
+ "0.6": 20,
78
+ "0.7": 13,
79
+ "0.8": 4,
80
+ "0.9": 0
81
+ },
82
+ "largest_dup_abs": {
83
+ "0.5": 4,
84
+ "0.6": 2,
85
+ "0.7": 2,
86
+ "0.8": 2,
87
+ "0.9": 1
88
+ },
89
+ "babel_mu1_abs": 0.37956738471984863,
90
+ "dup_excludes_dead": true
91
+ },
92
+ "seed1": {
93
+ "d_sae": 3072,
94
+ "d_in": 768,
95
+ "l0": 34,
96
+ "l0_raw": 34.20432915039063,
97
+ "recon_r2": 0.8978192650851368,
98
+ "l1_per_token": 255.75038515835834,
99
+ "stats_source": "stats",
100
+ "n_dead": 0,
101
+ "dead_rate": 0.0,
102
+ "alive_source": "stats",
103
+ "n_dense": 21,
104
+ "dense_rate": 0.0068359375,
105
+ "dense_threshold": 0.1,
106
+ "n_dup_signed": {
107
+ "0.5": 302,
108
+ "0.6": 0,
109
+ "0.7": 0,
110
+ "0.8": 0,
111
+ "0.9": 0
112
+ },
113
+ "n_dup_clusters_signed": {
114
+ "0.5": 146,
115
+ "0.6": 0,
116
+ "0.7": 0,
117
+ "0.8": 0,
118
+ "0.9": 0
119
+ },
120
+ "n_dup_removed_signed": {
121
+ "0.5": 156,
122
+ "0.6": 0,
123
+ "0.7": 0,
124
+ "0.8": 0,
125
+ "0.9": 0
126
+ },
127
+ "largest_dup_signed": {
128
+ "0.5": 4,
129
+ "0.6": 1,
130
+ "0.7": 1,
131
+ "0.8": 1,
132
+ "0.9": 1
133
+ },
134
+ "babel_mu1_signed": 0.37258124351501465,
135
+ "n_dup_abs": {
136
+ "0.5": 321,
137
+ "0.6": 36,
138
+ "0.7": 20,
139
+ "0.8": 4,
140
+ "0.9": 0
141
+ },
142
+ "n_dup_clusters_abs": {
143
+ "0.5": 155,
144
+ "0.6": 18,
145
+ "0.7": 10,
146
+ "0.8": 2,
147
+ "0.9": 0
148
+ },
149
+ "n_dup_removed_abs": {
150
+ "0.5": 166,
151
+ "0.6": 18,
152
+ "0.7": 10,
153
+ "0.8": 2,
154
+ "0.9": 0
155
+ },
156
+ "largest_dup_abs": {
157
+ "0.5": 4,
158
+ "0.6": 2,
159
+ "0.7": 2,
160
+ "0.8": 2,
161
+ "0.9": 1
162
+ },
163
+ "babel_mu1_abs": 0.3790799081325531,
164
+ "dup_excludes_dead": true
165
+ },
166
+ "seed2": {
167
+ "d_sae": 3072,
168
+ "d_in": 768,
169
+ "l0": 34,
170
+ "l0_raw": 34.06631484375,
171
+ "recon_r2": 0.8978286378898942,
172
+ "l1_per_token": 256.2946467344169,
173
+ "stats_source": "stats",
174
+ "n_dead": 0,
175
+ "dead_rate": 0.0,
176
+ "alive_source": "stats",
177
+ "n_dense": 20,
178
+ "dense_rate": 0.006510416666666667,
179
+ "dense_threshold": 0.1,
180
+ "n_dup_signed": {
181
+ "0.5": 288,
182
+ "0.6": 0,
183
+ "0.7": 0,
184
+ "0.8": 0,
185
+ "0.9": 0
186
+ },
187
+ "n_dup_clusters_signed": {
188
+ "0.5": 136,
189
+ "0.6": 0,
190
+ "0.7": 0,
191
+ "0.8": 0,
192
+ "0.9": 0
193
+ },
194
+ "n_dup_removed_signed": {
195
+ "0.5": 152,
196
+ "0.6": 0,
197
+ "0.7": 0,
198
+ "0.8": 0,
199
+ "0.9": 0
200
+ },
201
+ "largest_dup_signed": {
202
+ "0.5": 4,
203
+ "0.6": 1,
204
+ "0.7": 1,
205
+ "0.8": 1,
206
+ "0.9": 1
207
+ },
208
+ "babel_mu1_signed": 0.37247905135154724,
209
+ "n_dup_abs": {
210
+ "0.5": 305,
211
+ "0.6": 32,
212
+ "0.7": 18,
213
+ "0.8": 8,
214
+ "0.9": 0
215
+ },
216
+ "n_dup_clusters_abs": {
217
+ "0.5": 143,
218
+ "0.6": 16,
219
+ "0.7": 9,
220
+ "0.8": 4,
221
+ "0.9": 0
222
+ },
223
+ "n_dup_removed_abs": {
224
+ "0.5": 162,
225
+ "0.6": 16,
226
+ "0.7": 9,
227
+ "0.8": 4,
228
+ "0.9": 0
229
+ },
230
+ "largest_dup_abs": {
231
+ "0.5": 4,
232
+ "0.6": 2,
233
+ "0.7": 2,
234
+ "0.8": 2,
235
+ "0.9": 1
236
+ },
237
+ "babel_mu1_abs": 0.37893176078796387,
238
+ "dup_excludes_dead": true
239
+ }
240
+ },
241
+ "per_pair": {
242
+ "seed0\u2194seed1": {
243
+ "dcs_signed": 0.8927777409553528,
244
+ "dcs_abs": 0.8932483196258545,
245
+ "frac_sign_flipped": 0.009440104166666666,
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": 0.8908541202545166,
252
+ "0.6": 0.8927777409553528,
253
+ "0.7": 0.8927777409553528,
254
+ "0.8": 0.8927777409553528,
255
+ "0.9": 0.8927777409553528
256
+ },
257
+ "unique_frac_signed": {
258
+ "0.5": 0.9479166666666666,
259
+ "0.6": 1.0,
260
+ "0.7": 1.0,
261
+ "0.8": 1.0,
262
+ "0.9": 1.0
263
+ },
264
+ "dedup_dcs_signed": {
265
+ "0.5": 0.8444554681579272,
266
+ "0.6": 0.8927777409553528,
267
+ "0.7": 0.8927777409553528,
268
+ "0.8": 0.8927777409553528,
269
+ "0.9": 0.8927777409553528
270
+ },
271
+ "matched_frac_signed": {
272
+ "0.5": 0.8863932291666666,
273
+ "0.6": 0.9300130208333334,
274
+ "0.7": 0.8990885416666666,
275
+ "0.8": 0.8447265625,
276
+ "0.9": 0.75
277
+ },
278
+ "matched_cos_abs": {
279
+ "0.5": 0.8912460207939148,
280
+ "0.6": 0.8940423130989075,
281
+ "0.7": 0.8936648368835449,
282
+ "0.8": 0.8935339450836182,
283
+ "0.9": 0.8932483792304993
284
+ },
285
+ "unique_frac_abs": {
286
+ "0.5": 0.94482421875,
287
+ "0.6": 0.9938151041666666,
288
+ "0.7": 0.9962565104166666,
289
+ "0.8": 0.9990234375,
290
+ "0.9": 1.0
291
+ },
292
+ "dedup_dcs_abs": {
293
+ "0.5": 0.8420708253106568,
294
+ "0.6": 0.8885127545217983,
295
+ "0.7": 0.8903194118756801,
296
+ "0.8": 0.8926613533403724,
297
+ "0.9": 0.8932483792304993
298
+ },
299
+ "matched_frac_abs": {
300
+ "0.5": 0.8834635416666666,
301
+ "0.6": 0.9231770833333334,
302
+ "0.7": 0.8955078125,
303
+ "0.8": 0.8470052083333334,
304
+ "0.9": 0.75
305
+ }
306
+ },
307
+ "seed0\u2194seed2": {
308
+ "dcs_signed": 0.882415771484375,
309
+ "dcs_abs": 0.8828290104866028,
310
+ "frac_sign_flipped": 0.0078125,
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": 0.881562352180481,
317
+ "0.6": 0.882415771484375,
318
+ "0.7": 0.882415771484375,
319
+ "0.8": 0.882415771484375,
320
+ "0.9": 0.882415771484375
321
+ },
322
+ "unique_frac_signed": {
323
+ "0.5": 0.9485677083333334,
324
+ "0.6": 1.0,
325
+ "0.7": 1.0,
326
+ "0.8": 1.0,
327
+ "0.9": 1.0
328
+ },
329
+ "dedup_dcs_signed": {
330
+ "0.5": 0.8362215801607817,
331
+ "0.6": 0.882415771484375,
332
+ "0.7": 0.882415771484375,
333
+ "0.8": 0.882415771484375,
334
+ "0.9": 0.882415771484375
335
+ },
336
+ "matched_frac_signed": {
337
+ "0.5": 0.8798828125,
338
+ "0.6": 0.9108072916666666,
339
+ "0.7": 0.8782552083333334,
340
+ "0.8": 0.8317057291666666,
341
+ "0.9": 0.7233072916666666
342
+ },
343
+ "matched_cos_abs": {
344
+ "0.5": 0.8822004199028015,
345
+ "0.6": 0.8837887644767761,
346
+ "0.7": 0.8835625648498535,
347
+ "0.8": 0.8828677535057068,
348
+ "0.9": 0.8828290104866028
349
+ },
350
+ "unique_frac_abs": {
351
+ "0.5": 0.9454752604166666,
352
+ "0.6": 0.994140625,
353
+ "0.7": 0.9964192708333334,
354
+ "0.8": 0.9986979166666666,
355
+ "0.9": 1.0
356
+ },
357
+ "dedup_dcs_abs": {
358
+ "0.5": 0.8340986717472939,
359
+ "0.6": 0.87861031468492,
360
+ "0.7": 0.8803987666033208,
361
+ "0.8": 0.8817181861183295,
362
+ "0.9": 0.8828290104866028
363
+ },
364
+ "matched_frac_abs": {
365
+ "0.5": 0.876953125,
366
+ "0.6": 0.9049479166666666,
367
+ "0.7": 0.8753255208333334,
368
+ "0.8": 0.8304036458333334,
369
+ "0.9": 0.7233072916666666
370
+ }
371
+ },
372
+ "seed1\u2194seed2": {
373
+ "dcs_signed": 0.8868472576141357,
374
+ "dcs_abs": 0.8873586058616638,
375
+ "frac_sign_flipped": 0.008463541666666666,
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": 0.884628176689148,
382
+ "0.6": 0.8868473172187805,
383
+ "0.7": 0.8868473172187805,
384
+ "0.8": 0.8868473172187805,
385
+ "0.9": 0.8868473172187805
386
+ },
387
+ "unique_frac_signed": {
388
+ "0.5": 0.9498697916666666,
389
+ "0.6": 1.0,
390
+ "0.7": 1.0,
391
+ "0.8": 1.0,
392
+ "0.9": 1.0
393
+ },
394
+ "dedup_dcs_signed": {
395
+ "0.5": 0.8402815818941841,
396
+ "0.6": 0.8868473172187805,
397
+ "0.7": 0.8868473172187805,
398
+ "0.8": 0.8868473172187805,
399
+ "0.9": 0.8868473172187805
400
+ },
401
+ "matched_frac_signed": {
402
+ "0.5": 0.8815104166666666,
403
+ "0.6": 0.9163411458333334,
404
+ "0.7": 0.8857421875,
405
+ "0.8": 0.841796875,
406
+ "0.9": 0.7317708333333334
407
+ },
408
+ "matched_cos_abs": {
409
+ "0.5": 0.8856135010719299,
410
+ "0.6": 0.8882299661636353,
411
+ "0.7": 0.8876924514770508,
412
+ "0.8": 0.8877029418945312,
413
+ "0.9": 0.8873586058616638
414
+ },
415
+ "unique_frac_abs": {
416
+ "0.5": 0.9466145833333334,
417
+ "0.6": 0.9944661458333334,
418
+ "0.7": 0.9969075520833334,
419
+ "0.8": 0.9990234375,
420
+ "0.9": 1.0
421
+ },
422
+ "dedup_dcs_abs": {
423
+ "0.5": 0.8383346553115796,
424
+ "0.6": 0.8833146310644224,
425
+ "0.7": 0.8849473088048398,
426
+ "0.8": 0.8868360444903374,
427
+ "0.9": 0.8873586058616638
428
+ },
429
+ "matched_frac_abs": {
430
+ "0.5": 0.8785807291666666,
431
+ "0.6": 0.9108072916666666,
432
+ "0.7": 0.8834635416666666,
433
+ "0.8": 0.841796875,
434
+ "0.9": 0.7320963541666666
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": 20,
450
+ "n_dense_std": 0.5773502691896258,
451
+ "l0_mean": 34,
452
+ "l0_std": 0.07153421325772635,
453
+ "recon_r2_mean": 0.8977954543738922,
454
+ "recon_r2_std": 4.9580427810158145e-05,
455
+ "l1_per_token_mean": 255.69736072589058,
456
+ "l1_per_token_std": 0.6254861453644979,
457
+ "babel_mu1_signed_mean": 0.37275142470995587,
458
+ "babel_mu1_signed_std": 0.00038665450868511356,
459
+ "n_dup_signed_per_threshold": {
460
+ "0.5": 301.0,
461
+ "0.6": 0.0,
462
+ "0.7": 0.0,
463
+ "0.8": 0.0,
464
+ "0.9": 0.0
465
+ },
466
+ "n_dup_signed_band": 0.0,
467
+ "n_dup_signed_band_std": 0.0,
468
+ "n_dup_clusters_signed_per_threshold": {
469
+ "0.5": 143.66666666666666,
470
+ "0.6": 0.0,
471
+ "0.7": 0.0,
472
+ "0.8": 0.0,
473
+ "0.9": 0.0
474
+ },
475
+ "n_dup_clusters_signed_band": 0.0,
476
+ "n_dup_clusters_signed_band_std": 0.0,
477
+ "largest_dup_signed_per_threshold": {
478
+ "0.5": 4.0,
479
+ "0.6": 1.0,
480
+ "0.7": 1.0,
481
+ "0.8": 1.0,
482
+ "0.9": 1.0
483
+ },
484
+ "largest_dup_signed_band": 1.0,
485
+ "largest_dup_signed_band_std": 0.0,
486
+ "dcs_signed_mean": 0.8873469233512878,
487
+ "dcs_signed_std": 0.005199024159977689,
488
+ "dedup_dcs_signed_per_threshold": {
489
+ "0.5": 0.8403195434042976,
490
+ "0.6": 0.8873469432195028,
491
+ "0.7": 0.8873469432195028,
492
+ "0.8": 0.8873469432195028,
493
+ "0.9": 0.8873469432195028
494
+ },
495
+ "dedup_dcs_signed_band": 0.8873469432195028,
496
+ "dedup_dcs_signed_band_std": 0.005199021295861095,
497
+ "matched_cos_signed_per_threshold": {
498
+ "0.5": 0.8856815497080485,
499
+ "0.6": 0.8873469432195028,
500
+ "0.7": 0.8873469432195028,
501
+ "0.8": 0.8873469432195028,
502
+ "0.9": 0.8873469432195028
503
+ },
504
+ "matched_cos_signed_band": 0.8873469432195028,
505
+ "matched_cos_signed_band_std": 0.005199021295861095,
506
+ "unique_frac_signed_per_threshold": {
507
+ "0.5": 0.9487847222222222,
508
+ "0.6": 1.0,
509
+ "0.7": 1.0,
510
+ "0.8": 1.0,
511
+ "0.9": 1.0
512
+ },
513
+ "unique_frac_signed_band": 1.0,
514
+ "unique_frac_signed_band_std": 0.0,
515
+ "matched_frac_signed_per_threshold": {
516
+ "0.5": 0.882595486111111,
517
+ "0.6": 0.9190538194444445,
518
+ "0.7": 0.8876953125,
519
+ "0.8": 0.8394097222222222,
520
+ "0.9": 0.7350260416666666
521
+ },
522
+ "matched_frac_signed_band": 0.8452962239583334,
523
+ "matched_frac_signed_band_std": 0.010040813974395683,
524
+ "babel_mu1_abs_mean": 0.3791930178801219,
525
+ "babel_mu1_abs_std": 0.00033256549271153314,
526
+ "n_dup_abs_per_threshold": {
527
+ "0.5": 318.0,
528
+ "0.6": 36.0,
529
+ "0.7": 21.333333333333332,
530
+ "0.8": 6.666666666666667,
531
+ "0.9": 0.0
532
+ },
533
+ "n_dup_abs_band": 15.999999999999998,
534
+ "n_dup_abs_band_std": 2.179449471770337,
535
+ "n_dup_clusters_abs_per_threshold": {
536
+ "0.5": 151.0,
537
+ "0.6": 18.0,
538
+ "0.7": 10.666666666666666,
539
+ "0.8": 3.3333333333333335,
540
+ "0.9": 0.0
541
+ },
542
+ "n_dup_clusters_abs_band": 7.999999999999999,
543
+ "n_dup_clusters_abs_band_std": 1.0897247358851685,
544
+ "largest_dup_abs_per_threshold": {
545
+ "0.5": 4.0,
546
+ "0.6": 2.0,
547
+ "0.7": 2.0,
548
+ "0.8": 2.0,
549
+ "0.9": 1.0
550
+ },
551
+ "largest_dup_abs_band": 1.75,
552
+ "largest_dup_abs_band_std": 0.0,
553
+ "dcs_abs_mean": 0.8878119786580404,
554
+ "dcs_abs_std": 0.005224429241954978,
555
+ "dedup_dcs_abs_per_threshold": {
556
+ "0.5": 0.8381680507898434,
557
+ "0.6": 0.8834792334237136,
558
+ "0.7": 0.8852218290946136,
559
+ "0.8": 0.8870718613163464,
560
+ "0.9": 0.8878119985262553
561
+ },
562
+ "dedup_dcs_abs_band": 0.8858962305902321,
563
+ "dedup_dcs_abs_band_std": 0.005153995391320634,
564
+ "matched_cos_abs_per_threshold": {
565
+ "0.5": 0.8863533139228821,
566
+ "0.6": 0.888687014579773,
567
+ "0.7": 0.8883066177368164,
568
+ "0.8": 0.8880348801612854,
569
+ "0.9": 0.8878119985262553
570
+ },
571
+ "matched_cos_abs_band": 0.8882101277510325,
572
+ "matched_cos_abs_band_std": 0.005195743964577689,
573
+ "unique_frac_abs_per_threshold": {
574
+ "0.5": 0.9456380208333334,
575
+ "0.6": 0.994140625,
576
+ "0.7": 0.9965277777777778,
577
+ "0.8": 0.9989149305555555,
578
+ "0.9": 1.0
579
+ },
580
+ "unique_frac_abs_band": 0.9973958333333333,
581
+ "unique_frac_abs_band_std": 0.00017736405206471122,
582
+ "matched_frac_abs_per_threshold": {
583
+ "0.5": 0.879665798611111,
584
+ "0.6": 0.9129774305555555,
585
+ "0.7": 0.884765625,
586
+ "0.8": 0.8397352430555557,
587
+ "0.9": 0.735134548611111
588
+ },
589
+ "matched_frac_abs_band": 0.8431532118055556,
590
+ "matched_frac_abs_band_std": 0.01025853397418673,
591
+ "alive_source": "stats",
592
+ "frac_denominator": "d_sae",
593
+ "d_sae": 3072
594
+ }
595
+ }
lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 7.1,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 15,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }
lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_samples": 20000,
3
+ "n_positions": 20480000,
4
+ "l0_per_token_0": 34.167963525390626,
5
+ "l0_per_token_0.001": 34.167961572265625,
6
+ "l0_per_token_0.01": 34.167939111328124,
7
+ "l0_per_token_0.05": 34.167740185546876,
8
+ "l0_per_token_0.1": 34.16754931640625,
9
+ "l0_per_token_0.5": 34.164538037109374,
10
+ "l0_per_token_1.0": 34.15486186523437,
11
+ "l1_per_token": 255.04705028489641,
12
+ "r2_score": 0.8977384601466455,
13
+ "l2_loss": 2.3982820157353704
14
+ }
lm/gpt2-small_layer8/consdl/x4/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset": "fineweb",
3
+ "split": "train",
4
+ "hf_dataset": "HuggingFaceFW/fineweb",
5
+ "hf_dataset_config": null,
6
+ "hf_dataset_text_field": "text",
7
+ "context_length": 1024,
8
+ "num_samples_requested": 20000,
9
+ "num_samples_actual": 20000,
10
+ "batch_seqs": 32,
11
+ "stats_top_k": 100,
12
+ "stream_seed": 0,
13
+ "sparsity_thresholds": [
14
+ 0.001,
15
+ 0.01,
16
+ 0.05,
17
+ 0.1,
18
+ 0.5,
19
+ 1.0
20
+ ],
21
+ "model_hf_name": "openai-community/gpt2",
22
+ "hook_target_name": "transformer.h.8",
23
+ "tokenizer_hf_name": "openai-community/gpt2",
24
+ "d_sae": 3072,
25
+ "d_in": 768
26
+ }
lm/gpt2-small_layer8/consdl/x4/l0_32/seed1/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 4,
4
+ "architecture": "fista",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "fista_t_iters": 30,
9
+ "fista_step": 0.0,
10
+ "fista_lam": 7.1,
11
+ "fista_nonneg": true,
12
+ "fista_power_iters": 32,
13
+ "fista_r2_tol": 0.0001,
14
+ "fista_check_every": 3,
15
+ "fista_use_bf16": false,
16
+ "fista_step_cache_every": 50,
17
+ "fista_compile": false,
18
+ "fista_topk_k": 0,
19
+ "refit": true,
20
+ "refit_iters": 15,
21
+ "refit_thresh": 1e-08,
22
+ "class_type": "FistaSAE",
23
+ "module_name": "consdl.models.fista"
24
+ }