jjho-choi commited on
Commit
f43d1c5
·
verified ·
1 Parent(s): 80a2e72

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/batchtopk/x16/l0_128/code_metrics_step_final.json +136 -0
  2. lm/gpt2-small_layer8/batchtopk/x16/l0_128/dict_metrics_step_final.json +545 -0
  3. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/config.json +12 -0
  4. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/model.safetensors +3 -0
  5. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt +3 -0
  6. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt +3 -0
  7. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  8. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt +3 -0
  9. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt +3 -0
  10. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  11. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt +3 -0
  12. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/config.json +12 -0
  13. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/model.safetensors +3 -0
  14. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt +3 -0
  15. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt +3 -0
  16. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  17. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt +3 -0
  18. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt +3 -0
  19. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  20. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt +3 -0
  21. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/config.json +12 -0
  22. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/model.safetensors +3 -0
  23. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt +3 -0
  24. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt +3 -0
  25. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  26. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/max_act_per_token.pt +3 -0
  27. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/mean_act_per_token.pt +3 -0
  28. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json +26 -0
  29. lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/sparsity_per_token.pt +3 -0
  30. lm/gpt2-small_layer8/batchtopk/x16/l0_32/code_metrics_step_final.json +136 -0
  31. lm/gpt2-small_layer8/batchtopk/x16/l0_32/dict_metrics_step_final.json +545 -0
  32. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/config.json +12 -0
  33. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/model.safetensors +3 -0
  34. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt +3 -0
  35. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt +3 -0
  36. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  37. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt +3 -0
  38. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt +3 -0
  39. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json +26 -0
  40. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt +3 -0
  41. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/config.json +12 -0
  42. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/model.safetensors +3 -0
  43. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt +3 -0
  44. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt +3 -0
  45. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json +14 -0
  46. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt +3 -0
  47. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt +3 -0
  48. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/meta.json +26 -0
  49. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt +3 -0
  50. lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed2/config.json +12 -0
lm/gpt2-small_layer8/batchtopk/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": 127.60186833333333,
20
+ "recon_r2": 0.9549164696200511,
21
+ "l1_per_token": 607.733303906817
22
+ },
23
+ "seed1": {
24
+ "l0": 128.26901333333333,
25
+ "recon_r2": 0.9550364366052586,
26
+ "l1_per_token": 610.3533042558479
27
+ },
28
+ "seed2": {
29
+ "l0": 127.33541666666666,
30
+ "recon_r2": 0.9548217346630481,
31
+ "l1_per_token": 606.1776392972702
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7197823649387445,
38
+ "mcc_abs": 0.7198161701249538,
39
+ "act_iou": 0.3765977592859709,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9726636356267777,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8604219640832733,
44
+ "0.7": 0.8902659821617671,
45
+ "0.8": 0.9180785324351199,
46
+ "0.9": 0.9493570893912988
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8604668676010798,
50
+ "0.7": 0.8905839594546318,
51
+ "0.8": 0.9183884360757144,
52
+ "0.9": 0.9498763732181995
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.7164556125489968,
58
+ "mcc_abs": 0.7164774781849763,
59
+ "act_iou": 0.37359442046385594,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.9724552338792009,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8593533231762739,
64
+ "0.7": 0.8896758148365826,
65
+ "0.8": 0.9198472560480646,
66
+ "0.9": 0.9504334646407947
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8599169115633304,
70
+ "0.7": 0.890333888426092,
71
+ "0.8": 0.9201968236292594,
72
+ "0.9": 0.9505123240419678
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.7184330381742603,
78
+ "mcc_abs": 0.7184667161459387,
79
+ "act_iou": 0.3755397594742718,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.9727663797231461,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8577225733053204,
84
+ "0.7": 0.8891815496720875,
85
+ "0.8": 0.9174585580331756,
86
+ "0.9": 0.9486653999848231
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.8577840632335128,
90
+ "0.7": 0.8894980949335266,
91
+ "0.8": 0.9177318409381179,
92
+ "0.9": 0.9487187715656917
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": 127.73543277777777,
106
+ "l0_std": 0.48091608375848277,
107
+ "recon_r2_mean": 0.9549248802961192,
108
+ "recon_r2_std": 0.00010759779552195764,
109
+ "l1_per_token_mean": 608.088082486645,
110
+ "l1_per_token_std": 2.1103187773656535,
111
+ "input_r2_mean": 0.9726284164097082,
112
+ "input_r2_std": 0.00015853463025262014,
113
+ "mcc_signed_mean": 0.7182236718873338,
114
+ "mcc_signed_std": 0.00167322922734722,
115
+ "mcc_abs_mean": 0.718253454818623,
116
+ "mcc_abs_std": 0.0016795315605293117,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8591659535216225,
119
+ "0.7": 0.889707782223479,
120
+ "0.8": 0.9184614488387867,
121
+ "0.9": 0.9494853180056388
122
+ },
123
+ "matched_mcc_signed_band": 0.9042051256473818,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8593892807993077,
126
+ "0.7": 0.8901386476047501,
127
+ "0.8": 0.9187723668810305,
128
+ "0.9": 0.9497024896086197
129
+ },
130
+ "matched_mcc_abs_band": 0.904500696223427,
131
+ "act_iou_mean": 0.3752439797413662,
132
+ "act_iou_std": 0.0015233598583079414,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/dict_metrics_step_final.json ADDED
@@ -0,0 +1,545 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": 128,
22
+ "l0_raw": 127.7182478515625,
23
+ "recon_r2": 0.9548467507110822,
24
+ "l1_per_token": 607.7213934281276,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 242,
30
+ "dense_rate": 0.019694010416666668,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 470,
34
+ "0.6": 95,
35
+ "0.7": 21,
36
+ "0.8": 5,
37
+ "0.9": 2
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 232,
41
+ "0.6": 47,
42
+ "0.7": 10,
43
+ "0.8": 2,
44
+ "0.9": 1
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 3,
48
+ "0.6": 3,
49
+ "0.7": 3,
50
+ "0.8": 3,
51
+ "0.9": 2
52
+ },
53
+ "babel_mu1_signed": 0.3079283535480499,
54
+ "n_dup_abs": {
55
+ "0.5": 1342,
56
+ "0.6": 915,
57
+ "0.7": 807,
58
+ "0.8": 751,
59
+ "0.9": 656
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 667,
63
+ "0.6": 456,
64
+ "0.7": 403,
65
+ "0.8": 375,
66
+ "0.9": 328
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 3,
70
+ "0.6": 3,
71
+ "0.7": 3,
72
+ "0.8": 3,
73
+ "0.9": 2
74
+ },
75
+ "babel_mu1_abs": 0.36509886384010315
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 128,
81
+ "l0_raw": 128.3524513671875,
82
+ "recon_r2": 0.9549618942987953,
83
+ "l1_per_token": 610.2288667752479,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 238,
89
+ "dense_rate": 0.019368489583333332,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 497,
93
+ "0.6": 95,
94
+ "0.7": 19,
95
+ "0.8": 5,
96
+ "0.9": 3
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 245,
100
+ "0.6": 46,
101
+ "0.7": 9,
102
+ "0.8": 2,
103
+ "0.9": 1
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 3,
107
+ "0.6": 3,
108
+ "0.7": 3,
109
+ "0.8": 3,
110
+ "0.9": 3
111
+ },
112
+ "babel_mu1_signed": 0.30881133675575256,
113
+ "n_dup_abs": {
114
+ "0.5": 1375,
115
+ "0.6": 928,
116
+ "0.7": 807,
117
+ "0.8": 739,
118
+ "0.9": 631
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 681,
122
+ "0.6": 461,
123
+ "0.7": 402,
124
+ "0.8": 369,
125
+ "0.9": 315
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 3,
129
+ "0.6": 3,
130
+ "0.7": 3,
131
+ "0.8": 3,
132
+ "0.9": 3
133
+ },
134
+ "babel_mu1_abs": 0.3659749925136566
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 127,
140
+ "l0_raw": 127.48565991210937,
141
+ "recon_r2": 0.954758292979353,
142
+ "l1_per_token": 606.2793811777946,
143
+ "stats_source": "stats",
144
+ "n_dead": 0,
145
+ "dead_rate": 0.0,
146
+ "alive_source": "stats",
147
+ "n_dense": 238,
148
+ "dense_rate": 0.019368489583333332,
149
+ "dense_threshold": 0.1,
150
+ "n_dup_signed": {
151
+ "0.5": 532,
152
+ "0.6": 123,
153
+ "0.7": 31,
154
+ "0.8": 9,
155
+ "0.9": 2
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 261,
159
+ "0.6": 61,
160
+ "0.7": 15,
161
+ "0.8": 4,
162
+ "0.9": 1
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 4,
166
+ "0.6": 3,
167
+ "0.7": 3,
168
+ "0.8": 3,
169
+ "0.9": 2
170
+ },
171
+ "babel_mu1_signed": 0.30918964743614197,
172
+ "n_dup_abs": {
173
+ "0.5": 1405,
174
+ "0.6": 943,
175
+ "0.7": 785,
176
+ "0.8": 713,
177
+ "0.9": 638
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 694,
181
+ "0.6": 469,
182
+ "0.7": 390,
183
+ "0.8": 356,
184
+ "0.9": 319
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 4,
188
+ "0.6": 3,
189
+ "0.7": 3,
190
+ "0.8": 3,
191
+ "0.9": 2
192
+ },
193
+ "babel_mu1_abs": 0.36544737219810486
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.7136623859405518,
199
+ "dcs_abs": 0.7151963114738464,
200
+ "frac_sign_flipped": 0.022786458333333332,
201
+ "alive_source": "stats",
202
+ "n_alive_a": 12288,
203
+ "n_alive_b": 12288,
204
+ "matched_cos_signed": {
205
+ "0.5": 0.7085920572280884,
206
+ "0.6": 0.7129131555557251,
207
+ "0.7": 0.7134965658187866,
208
+ "0.8": 0.7136114835739136,
209
+ "0.9": 0.7136274576187134
210
+ },
211
+ "unique_frac_signed": {
212
+ "0.5": 0.9800618489583334,
213
+ "0.6": 0.9960530598958334,
214
+ "0.7": 0.9991455078125,
215
+ "0.8": 0.999755859375,
216
+ "0.9": 0.9998779296875
217
+ },
218
+ "dedup_dcs_signed": {
219
+ "0.5": 0.6944640417641494,
220
+ "0.6": 0.7100993300312742,
221
+ "0.7": 0.7128868885774864,
222
+ "0.8": 0.7134372620203067,
223
+ "0.9": 0.7135403448919533
224
+ },
225
+ "matched_frac_signed": {
226
+ "0.5": 0.7732747395833334,
227
+ "0.6": 0.7207845052083334,
228
+ "0.7": 0.6289876302083334,
229
+ "0.8": 0.5049641927083334,
230
+ "0.9": 0.2537434895833333
231
+ },
232
+ "matched_cos_abs": {
233
+ "0.5": 0.7161705493927002,
234
+ "0.6": 0.7201769948005676,
235
+ "0.7": 0.7203535437583923,
236
+ "0.8": 0.7201673984527588,
237
+ "0.9": 0.7195829153060913
238
+ },
239
+ "unique_frac_abs": {
240
+ "0.5": 0.9442952473958334,
241
+ "0.6": 0.9623209635416666,
242
+ "0.7": 0.9670817057291666,
243
+ "0.8": 0.9696451822916666,
244
+ "0.9": 0.9737955729166666
245
+ },
246
+ "dedup_dcs_abs": {
247
+ "0.5": 0.6762764461163897,
248
+ "0.6": 0.693041419557024,
249
+ "0.7": 0.696640733825916,
250
+ "0.8": 0.6983068483532406,
251
+ "0.9": 0.7007266572715404
252
+ },
253
+ "matched_frac_abs": {
254
+ "0.5": 0.7564290364583334,
255
+ "0.6": 0.7080078125,
256
+ "0.7": 0.6189778645833334,
257
+ "0.8": 0.4969889322916667,
258
+ "0.9": 0.24943033854166666
259
+ }
260
+ },
261
+ "seed0\u2194seed2": {
262
+ "dcs_signed": 0.7091007828712463,
263
+ "dcs_abs": 0.7105391621589661,
264
+ "frac_sign_flipped": 0.025797526041666668,
265
+ "alive_source": "stats",
266
+ "n_alive_a": 12288,
267
+ "n_alive_b": 12288,
268
+ "matched_cos_signed": {
269
+ "0.5": 0.7051715850830078,
270
+ "0.6": 0.7086870670318604,
271
+ "0.7": 0.7091602683067322,
272
+ "0.8": 0.7090668082237244,
273
+ "0.9": 0.7090792059898376
274
+ },
275
+ "unique_frac_signed": {
276
+ "0.5": 0.9792887369791666,
277
+ "0.6": 0.9955240885416666,
278
+ "0.7": 0.9989013671875,
279
+ "0.8": 0.9996744791666666,
280
+ "0.9": 0.9999186197916666
281
+ },
282
+ "dedup_dcs_signed": {
283
+ "0.5": 0.6905665909095356,
284
+ "0.6": 0.7055150464681598,
285
+ "0.7": 0.7083811615666491,
286
+ "0.8": 0.7088359922054224,
287
+ "0.9": 0.7090215009763293
288
+ },
289
+ "matched_frac_signed": {
290
+ "0.5": 0.7677408854166666,
291
+ "0.6": 0.7119140625,
292
+ "0.7": 0.6239420572916666,
293
+ "0.8": 0.4934895833333333,
294
+ "0.9": 0.2490234375
295
+ },
296
+ "matched_cos_abs": {
297
+ "0.5": 0.7125113606452942,
298
+ "0.6": 0.7156597971916199,
299
+ "0.7": 0.7157593965530396,
300
+ "0.8": 0.7155851125717163,
301
+ "0.9": 0.715090811252594
302
+ },
303
+ "unique_frac_abs": {
304
+ "0.5": 0.943603515625,
305
+ "0.6": 0.9620361328125,
306
+ "0.7": 0.9674886067708334,
307
+ "0.8": 0.9701741536458334,
308
+ "0.9": 0.9736735026041666
309
+ },
310
+ "dedup_dcs_abs": {
311
+ "0.5": 0.6723282248276519,
312
+ "0.6": 0.688490583699604,
313
+ "0.7": 0.6924890613542327,
314
+ "0.8": 0.6942421809508232,
315
+ "0.9": 0.6962649748723682
316
+ },
317
+ "matched_frac_abs": {
318
+ "0.5": 0.750244140625,
319
+ "0.6": 0.6988118489583334,
320
+ "0.7": 0.614501953125,
321
+ "0.8": 0.486572265625,
322
+ "0.9": 0.2451171875
323
+ }
324
+ },
325
+ "seed1\u2194seed2": {
326
+ "dcs_signed": 0.7128297686576843,
327
+ "dcs_abs": 0.7144290804862976,
328
+ "frac_sign_flipped": 0.027262369791666668,
329
+ "alive_source": "stats",
330
+ "n_alive_a": 12288,
331
+ "n_alive_b": 12288,
332
+ "matched_cos_signed": {
333
+ "0.5": 0.7082353830337524,
334
+ "0.6": 0.7123907804489136,
335
+ "0.7": 0.7128813862800598,
336
+ "0.8": 0.7128597497940063,
337
+ "0.9": 0.7127940654754639
338
+ },
339
+ "unique_frac_signed": {
340
+ "0.5": 0.9787190755208334,
341
+ "0.6": 0.9954833984375,
342
+ "0.7": 0.9989420572916666,
343
+ "0.8": 0.9996744791666666,
344
+ "0.9": 0.9998779296875
345
+ },
346
+ "dedup_dcs_signed": {
347
+ "0.5": 0.6931634793339375,
348
+ "0.6": 0.7091731951368274,
349
+ "0.7": 0.7121271986155382,
350
+ "0.8": 0.7126276990942035,
351
+ "0.9": 0.7127070544811431
352
+ },
353
+ "matched_frac_signed": {
354
+ "0.5": 0.7713216145833334,
355
+ "0.6": 0.7215983072916666,
356
+ "0.7": 0.6263020833333334,
357
+ "0.8": 0.503173828125,
358
+ "0.9": 0.2534993489583333
359
+ },
360
+ "matched_cos_abs": {
361
+ "0.5": 0.7156670689582825,
362
+ "0.6": 0.719334065914154,
363
+ "0.7": 0.7195640206336975,
364
+ "0.8": 0.7192935943603516,
365
+ "0.9": 0.7184290885925293
366
+ },
367
+ "unique_frac_abs": {
368
+ "0.5": 0.9428304036458334,
369
+ "0.6": 0.9617106119791666,
370
+ "0.7": 0.9674479166666666,
371
+ "0.8": 0.9704182942708334,
372
+ "0.9": 0.9741617838541666
373
+ },
374
+ "dedup_dcs_abs": {
375
+ "0.5": 0.6747526715019679,
376
+ "0.6": 0.6917912047477632,
377
+ "0.7": 0.6961407126703609,
378
+ "0.8": 0.6980156629191091,
379
+ "0.9": 0.6998661625160214
380
+ },
381
+ "matched_frac_abs": {
382
+ "0.5": 0.7534993489583334,
383
+ "0.6": 0.70947265625,
384
+ "0.7": 0.616943359375,
385
+ "0.8": 0.4955240885416667,
386
+ "0.9": 0.24951171875
387
+ }
388
+ }
389
+ },
390
+ "summary": {
391
+ "n_seeds": 3,
392
+ "n_pairs": 3,
393
+ "report_band": [
394
+ 0.6,
395
+ 0.7,
396
+ 0.8,
397
+ 0.9
398
+ ],
399
+ "n_dead_mean": 0,
400
+ "n_dead_std": 0.0,
401
+ "n_dense_mean": 239,
402
+ "n_dense_std": 2.3094010767585034,
403
+ "l0_mean": 128,
404
+ "l0_std": 0.4486430162394102,
405
+ "recon_r2_mean": 0.95485564599641,
406
+ "recon_r2_std": null,
407
+ "l1_per_token_mean": 608.0765471270566,
408
+ "l1_per_token_std": 1.9985449960435793,
409
+ "babel_mu1_signed_mean": 0.3086431125799815,
410
+ "babel_mu1_signed_std": 0.00064725582115573,
411
+ "n_dup_signed_per_threshold": {
412
+ "0.5": 499.6666666666667,
413
+ "0.6": 104.33333333333333,
414
+ "0.7": 23.666666666666668,
415
+ "0.8": 6.333333333333333,
416
+ "0.9": 2.3333333333333335
417
+ },
418
+ "n_dup_signed_band": 34.16666666666667,
419
+ "n_dup_signed_band_std": 6.135620044733321,
420
+ "n_dup_clusters_signed_per_threshold": {
421
+ "0.5": 246.0,
422
+ "0.6": 51.333333333333336,
423
+ "0.7": 11.333333333333334,
424
+ "0.8": 2.6666666666666665,
425
+ "0.9": 1.0
426
+ },
427
+ "n_dup_clusters_signed_band": 16.583333333333336,
428
+ "n_dup_clusters_signed_band_std": 3.185252475602732,
429
+ "largest_dup_signed_per_threshold": {
430
+ "0.5": 3.3333333333333335,
431
+ "0.6": 3.0,
432
+ "0.7": 3.0,
433
+ "0.8": 3.0,
434
+ "0.9": 2.3333333333333335
435
+ },
436
+ "largest_dup_signed_band": 2.8333333333333335,
437
+ "largest_dup_signed_band_std": 0.14433756729740643,
438
+ "dcs_signed_mean": 0.7118643124898275,
439
+ "dcs_signed_std": 0.0024292251541414983,
440
+ "dedup_dcs_signed_per_threshold": {
441
+ "0.5": 0.6927313706692075,
442
+ "0.6": 0.7082625238787538,
443
+ "0.7": 0.7111317495865578,
444
+ "0.8": 0.7116336511066441,
445
+ "0.9": 0.7117563001164752
446
+ },
447
+ "dedup_dcs_signed_band": 0.7106960561721076,
448
+ "dedup_dcs_signed_band_std": 0.0024241539848505723,
449
+ "matched_cos_signed_per_threshold": {
450
+ "0.5": 0.7073330084482828,
451
+ "0.6": 0.7113303343454996,
452
+ "0.7": 0.7118460734685262,
453
+ "0.8": 0.7118460138638815,
454
+ "0.9": 0.7118335763613383
455
+ },
456
+ "matched_cos_signed_band": 0.7117139995098114,
457
+ "matched_cos_signed_band_std": 0.0023763298445997004,
458
+ "unique_frac_signed_per_threshold": {
459
+ "0.5": 0.9793565538194445,
460
+ "0.6": 0.9956868489583334,
461
+ "0.7": 0.9989963107638888,
462
+ "0.8": 0.9997016059027777,
463
+ "0.9": 0.9998914930555555
464
+ },
465
+ "unique_frac_signed_band": 0.9985690646701388,
466
+ "unique_frac_signed_band_std": 0.00012050615535432782,
467
+ "matched_frac_signed_per_threshold": {
468
+ "0.5": 0.7707790798611112,
469
+ "0.6": 0.7180989583333334,
470
+ "0.7": 0.6264105902777778,
471
+ "0.8": 0.5005425347222222,
472
+ "0.9": 0.2520887586805555
473
+ },
474
+ "matched_frac_signed_band": 0.5242852105034722,
475
+ "matched_frac_signed_band_std": 0.0040934190861738485,
476
+ "babel_mu1_abs_mean": 0.3655070761839549,
477
+ "babel_mu1_abs_std": 0.00044110518881602777,
478
+ "n_dup_abs_per_threshold": {
479
+ "0.5": 1374.0,
480
+ "0.6": 928.6666666666666,
481
+ "0.7": 799.6666666666666,
482
+ "0.8": 734.3333333333334,
483
+ "0.9": 641.6666666666666
484
+ },
485
+ "n_dup_abs_band": 776.0833333333333,
486
+ "n_dup_abs_band_std": 6.251666444503684,
487
+ "n_dup_clusters_abs_per_threshold": {
488
+ "0.5": 680.6666666666666,
489
+ "0.6": 462.0,
490
+ "0.7": 398.3333333333333,
491
+ "0.8": 366.6666666666667,
492
+ "0.9": 320.6666666666667
493
+ },
494
+ "n_dup_clusters_abs_band": 386.9166666666667,
495
+ "n_dup_clusters_abs_band_std": 3.50297492616395,
496
+ "largest_dup_abs_per_threshold": {
497
+ "0.5": 3.3333333333333335,
498
+ "0.6": 3.0,
499
+ "0.7": 3.0,
500
+ "0.8": 3.0,
501
+ "0.9": 2.3333333333333335
502
+ },
503
+ "largest_dup_abs_band": 2.8333333333333335,
504
+ "largest_dup_abs_band_std": 0.14433756729740643,
505
+ "dcs_abs_mean": 0.71338818470637,
506
+ "dcs_abs_std": 0.0024969697542773614,
507
+ "dedup_dcs_abs_per_threshold": {
508
+ "0.5": 0.6744524474820032,
509
+ "0.6": 0.6911077360014638,
510
+ "0.7": 0.6950901692835032,
511
+ "0.8": 0.6968548974077243,
512
+ "0.9": 0.6989525982199766
513
+ },
514
+ "dedup_dcs_abs_band": 0.695501350228167,
515
+ "dedup_dcs_abs_band_std": 0.0023060516767600663,
516
+ "matched_cos_abs_per_threshold": {
517
+ "0.5": 0.7147829929987589,
518
+ "0.6": 0.7183902859687805,
519
+ "0.7": 0.7185589869817098,
520
+ "0.8": 0.7183487017949423,
521
+ "0.9": 0.7177009383837382
522
+ },
523
+ "matched_cos_abs_band": 0.7182497282822927,
524
+ "matched_cos_abs_band_std": 0.002404664994481575,
525
+ "unique_frac_abs_per_threshold": {
526
+ "0.5": 0.943576388888889,
527
+ "0.6": 0.9620225694444443,
528
+ "0.7": 0.9673394097222222,
529
+ "0.8": 0.9700792100694445,
530
+ "0.9": 0.973876953125
531
+ },
532
+ "unique_frac_abs_band": 0.9683295355902778,
533
+ "unique_frac_abs_band_std": 0.0001125126140885089,
534
+ "matched_frac_abs_per_threshold": {
535
+ "0.5": 0.753390842013889,
536
+ "0.6": 0.7054307725694445,
537
+ "0.7": 0.6168077256944445,
538
+ "0.8": 0.4930284288194445,
539
+ "0.9": 0.24801974826388887
540
+ },
541
+ "matched_frac_abs_band": 0.5158216688368056,
542
+ "matched_frac_abs_band_std": 0.003965998166427926,
543
+ "alive_source": "stats"
544
+ }
545
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "top_k": 128,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4542156a37ee3ec99e7ba415d096952c436b8464c17fed4820dd34d93c322549
3
+ size 75550084
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e903fd90d7211921b1f0515505b4937f0b2d8dd8b97c858206a70538901b7e5
3
+ size 99648
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45e5038b4f48b41b4f34484b212a519f5c329d0966db5b0bfc51e8a2697abca7
3
+ size 50471
lm/gpt2-small_layer8/batchtopk/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": 127.7182478515625,
5
+ "l0_per_token_0.001": 127.7182478515625,
6
+ "l0_per_token_0.01": 127.7182478515625,
7
+ "l0_per_token_0.05": 127.7182478515625,
8
+ "l0_per_token_0.1": 127.7182478515625,
9
+ "l0_per_token_0.5": 127.7182478515625,
10
+ "l0_per_token_1.0": 127.7182478515625,
11
+ "l1_per_token": 607.7213934281276,
12
+ "r2_score": 0.9548467507110822,
13
+ "l2_loss": 1.058953591401838
14
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ea7efe657c3b33aa7815ec8560dcf7665dde20243847cb0588d0c262a5dec73
3
+ size 50382
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c339d6727326b9cbe3da1f658fed6393fe63b7604c0c5b5ca0728ca54588955
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4e75a4e30e0d7f1bc6b5b245362a9956cabe7ebddc5f93699cc42a61da926af
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "top_k": 128,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c808be8f317e843dfd2f0bcba803f04b3317529ddb4d8f03dc89c0445b3ba893
3
+ size 75550084
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f729fb21572179aef4947eba759e49089330dc258139f1b0fb4f193c9db06960
3
+ size 99648
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57b51963dd53853f0f9514151a761dffc57ee52c9858e28ae401913c0a435387
3
+ size 50471
lm/gpt2-small_layer8/batchtopk/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": 128.3524513671875,
5
+ "l0_per_token_0.001": 128.3524513671875,
6
+ "l0_per_token_0.01": 128.3524513671875,
7
+ "l0_per_token_0.05": 128.3524513671875,
8
+ "l0_per_token_0.1": 128.3524513671875,
9
+ "l0_per_token_0.5": 128.3524513671875,
10
+ "l0_per_token_1.0": 128.3524513671875,
11
+ "l1_per_token": 610.2288667752479,
12
+ "r2_score": 0.9549618942987953,
13
+ "l2_loss": 1.0562531940294237
14
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9edd29217b79ee3dee47e88e274db68ffb05cdc1dc6ea9d33894806ce7eead3d
3
+ size 50382
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fde69c672847047db110cadad3443396db29fcd9613828ff9f5bda51a2914b8
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f268691c6c89f18ce1e2c30a68f361e818c37455ff20a39ad770ec1eb161695f
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "top_k": 128,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81d2f343f0dbeda4e2ffa6d049714ebc2c2f45a8f8ecbffa58115ccb8c7f4dac
3
+ size 75550084
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:367f01c34387df42f6df2e31a1ed498cf7936b9e4ea04da1e2cfc1f56fc4938f
3
+ size 99648
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:78e83f048ecef584c406203f717dd5d89ea771421359414662583767ec50c008
3
+ size 50471
lm/gpt2-small_layer8/batchtopk/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": 127.48565991210937,
5
+ "l0_per_token_0.001": 127.48565991210937,
6
+ "l0_per_token_0.01": 127.48565991210937,
7
+ "l0_per_token_0.05": 127.48565991210937,
8
+ "l0_per_token_0.1": 127.48565991210937,
9
+ "l0_per_token_0.5": 127.48565991210937,
10
+ "l0_per_token_1.0": 127.48565991210937,
11
+ "l1_per_token": 606.2793811777946,
12
+ "r2_score": 0.954758292979353,
13
+ "l2_loss": 1.061028140502446
14
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/max_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d02e06c8cfeed8e8ce03a03f474e41cd868e20c1b45dc50467599f5fc79fa9b8
3
+ size 50382
lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/mean_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cca20f1e17748b6887b333f51da83fe12658a498e6a1a0a41dc8234f91be04ac
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/sparsity_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61af4facba39ec6d7e84c402db4b1cfc7248324f7ac99a2476e09b071008d3b8
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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": 31.919266666666665,
20
+ "recon_r2": 0.9193748679709128,
21
+ "l1_per_token": 304.20424793829403
22
+ },
23
+ "seed1": {
24
+ "l0": 31.818753333333333,
25
+ "recon_r2": 0.9194302259933855,
26
+ "l1_per_token": 306.4840420251751
27
+ },
28
+ "seed2": {
29
+ "l0": 31.798905,
30
+ "recon_r2": 0.9193767591649853,
31
+ "l1_per_token": 305.6252251752166
32
+ }
33
+ },
34
+ "per_pair": {
35
+ "seed0\u2194seed1": {
36
+ "alive_source": "stats",
37
+ "mcc_signed": 0.7400087756930157,
38
+ "mcc_abs": 0.7400210476221692,
39
+ "act_iou": 0.43364045041840676,
40
+ "act_iou_n_pairs": 12288,
41
+ "input_r2": 0.9686586865788722,
42
+ "matched_mcc_signed": {
43
+ "0.6": 0.8444406047600785,
44
+ "0.7": 0.8733211255663039,
45
+ "0.8": 0.9007055349628019,
46
+ "0.9": 0.9319587076178429
47
+ },
48
+ "matched_mcc_abs": {
49
+ "0.6": 0.8424924662374571,
50
+ "0.7": 0.8729844102628117,
51
+ "0.8": 0.9003978364915014,
52
+ "0.9": 0.9317171967888204
53
+ }
54
+ },
55
+ "seed0\u2194seed2": {
56
+ "alive_source": "stats",
57
+ "mcc_signed": 0.744693272009517,
58
+ "mcc_abs": 0.74470590875556,
59
+ "act_iou": 0.43729593880994333,
60
+ "act_iou_n_pairs": 12288,
61
+ "input_r2": 0.9695608845428498,
62
+ "matched_mcc_signed": {
63
+ "0.6": 0.8443687195173901,
64
+ "0.7": 0.8749583129128411,
65
+ "0.8": 0.9012890242036748,
66
+ "0.9": 0.9319492239243443
67
+ },
68
+ "matched_mcc_abs": {
69
+ "0.6": 0.8433518751414391,
70
+ "0.7": 0.8743328909963229,
71
+ "0.8": 0.9008074760540641,
72
+ "0.9": 0.9315127755633315
73
+ }
74
+ },
75
+ "seed1\u2194seed2": {
76
+ "alive_source": "stats",
77
+ "mcc_signed": 0.742542751517254,
78
+ "mcc_abs": 0.742554242294822,
79
+ "act_iou": 0.4358845957507067,
80
+ "act_iou_n_pairs": 12288,
81
+ "input_r2": 0.9696325740204559,
82
+ "matched_mcc_signed": {
83
+ "0.6": 0.8449652732784033,
84
+ "0.7": 0.8733104171762524,
85
+ "0.8": 0.9015445092077589,
86
+ "0.9": 0.9318724590191894
87
+ },
88
+ "matched_mcc_abs": {
89
+ "0.6": 0.844412736067978,
90
+ "0.7": 0.8730420062902585,
91
+ "0.8": 0.9014987154182973,
92
+ "0.9": 0.9316214859387513
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": 31.845641666666666,
106
+ "l0_std": 0.06452882730049893,
107
+ "recon_r2_mean": 0.9193939510430945,
108
+ "recon_r2_std": 3.142925655682355e-05,
109
+ "l1_per_token_mean": 305.4378383795619,
110
+ "l1_per_token_std": 1.1513907364720959,
111
+ "input_r2_mean": 0.969284048380726,
112
+ "input_r2_std": 0.0005427641133247415,
113
+ "mcc_signed_mean": 0.7424149330732622,
114
+ "mcc_signed_std": 0.0023448623841916204,
115
+ "mcc_abs_mean": 0.7424270662241838,
116
+ "mcc_abs_std": 0.002345018395342776,
117
+ "matched_mcc_signed_per_threshold": {
118
+ "0.6": 0.8445915325186241,
119
+ "0.7": 0.8738632852184658,
120
+ "0.8": 0.9011796894580786,
121
+ "0.9": 0.9319267968537922
122
+ },
123
+ "matched_mcc_signed_band": 0.8878903260122402,
124
+ "matched_mcc_abs_per_threshold": {
125
+ "0.6": 0.8434190258156247,
126
+ "0.7": 0.8734531025164644,
127
+ "0.8": 0.9009013426546209,
128
+ "0.9": 0.9316171527636343
129
+ },
130
+ "matched_mcc_abs_band": 0.887347655937586,
131
+ "act_iou_mean": 0.43560699499301897,
132
+ "act_iou_std": 0.0018434873150272191,
133
+ "act_iou_n_pairs": 12288.0,
134
+ "alive_source": "stats"
135
+ }
136
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/dict_metrics_step_final.json ADDED
@@ -0,0 +1,545 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": 2,
3
+ "step": "final",
4
+ "linkage": "complete",
5
+ "thresholds": [
6
+ 0.5,
7
+ 0.6,
8
+ 0.7,
9
+ 0.8,
10
+ 0.9
11
+ ],
12
+ "seeds": [
13
+ "seed0",
14
+ "seed1",
15
+ "seed2"
16
+ ],
17
+ "per_seed": {
18
+ "seed0": {
19
+ "d_sae": 12288,
20
+ "d_in": 768,
21
+ "l0": 32,
22
+ "l0_raw": 31.96863642578125,
23
+ "recon_r2": 0.9192533434187236,
24
+ "l1_per_token": 304.34399717607556,
25
+ "stats_source": "stats",
26
+ "n_dead": 0,
27
+ "dead_rate": 0.0,
28
+ "alive_source": "stats",
29
+ "n_dense": 20,
30
+ "dense_rate": 0.0016276041666666667,
31
+ "dense_threshold": 0.1,
32
+ "n_dup_signed": {
33
+ "0.5": 1951,
34
+ "0.6": 660,
35
+ "0.7": 227,
36
+ "0.8": 138,
37
+ "0.9": 120
38
+ },
39
+ "n_dup_clusters_signed": {
40
+ "0.5": 877,
41
+ "0.6": 260,
42
+ "0.7": 51,
43
+ "0.8": 8,
44
+ "0.9": 7
45
+ },
46
+ "largest_dup_signed": {
47
+ "0.5": 128,
48
+ "0.6": 127,
49
+ "0.7": 127,
50
+ "0.8": 99,
51
+ "0.9": 51
52
+ },
53
+ "babel_mu1_signed": 0.40909543633461,
54
+ "n_dup_abs": {
55
+ "0.5": 2123,
56
+ "0.6": 801,
57
+ "0.7": 340,
58
+ "0.8": 230,
59
+ "0.9": 180
60
+ },
61
+ "n_dup_clusters_abs": {
62
+ "0.5": 959,
63
+ "0.6": 329,
64
+ "0.7": 107,
65
+ "0.8": 54,
66
+ "0.9": 37
67
+ },
68
+ "largest_dup_abs": {
69
+ "0.5": 128,
70
+ "0.6": 127,
71
+ "0.7": 127,
72
+ "0.8": 99,
73
+ "0.9": 51
74
+ },
75
+ "babel_mu1_abs": 0.41840028762817383
76
+ },
77
+ "seed1": {
78
+ "d_sae": 12288,
79
+ "d_in": 768,
80
+ "l0": 32,
81
+ "l0_raw": 31.867031298828124,
82
+ "recon_r2": 0.9193055911896232,
83
+ "l1_per_token": 306.4091118377043,
84
+ "stats_source": "stats",
85
+ "n_dead": 0,
86
+ "dead_rate": 0.0,
87
+ "alive_source": "stats",
88
+ "n_dense": 22,
89
+ "dense_rate": 0.0017903645833333333,
90
+ "dense_threshold": 0.1,
91
+ "n_dup_signed": {
92
+ "0.5": 1930,
93
+ "0.6": 651,
94
+ "0.7": 240,
95
+ "0.8": 138,
96
+ "0.9": 122
97
+ },
98
+ "n_dup_clusters_signed": {
99
+ "0.5": 871,
100
+ "0.6": 260,
101
+ "0.7": 59,
102
+ "0.8": 11,
103
+ "0.9": 22
104
+ },
105
+ "largest_dup_signed": {
106
+ "0.5": 124,
107
+ "0.6": 124,
108
+ "0.7": 80,
109
+ "0.8": 44,
110
+ "0.9": 22
111
+ },
112
+ "babel_mu1_signed": 0.4086414575576782,
113
+ "n_dup_abs": {
114
+ "0.5": 2112,
115
+ "0.6": 796,
116
+ "0.7": 352,
117
+ "0.8": 232,
118
+ "0.9": 189
119
+ },
120
+ "n_dup_clusters_abs": {
121
+ "0.5": 957,
122
+ "0.6": 331,
123
+ "0.7": 114,
124
+ "0.8": 57,
125
+ "0.9": 55
126
+ },
127
+ "largest_dup_abs": {
128
+ "0.5": 124,
129
+ "0.6": 124,
130
+ "0.7": 80,
131
+ "0.8": 44,
132
+ "0.9": 22
133
+ },
134
+ "babel_mu1_abs": 0.4184608459472656
135
+ },
136
+ "seed2": {
137
+ "d_sae": 12288,
138
+ "d_in": 768,
139
+ "l0": 32,
140
+ "l0_raw": 31.848321533203126,
141
+ "recon_r2": 0.9192398937198489,
142
+ "l1_per_token": 305.4881751320176,
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": 1969,
152
+ "0.6": 635,
153
+ "0.7": 243,
154
+ "0.8": 145,
155
+ "0.9": 132
156
+ },
157
+ "n_dup_clusters_signed": {
158
+ "0.5": 882,
159
+ "0.6": 245,
160
+ "0.7": 55,
161
+ "0.8": 7,
162
+ "0.9": 9
163
+ },
164
+ "largest_dup_signed": {
165
+ "0.5": 134,
166
+ "0.6": 134,
167
+ "0.7": 133,
168
+ "0.8": 133,
169
+ "0.9": 63
170
+ },
171
+ "babel_mu1_signed": 0.40897050499916077,
172
+ "n_dup_abs": {
173
+ "0.5": 2138,
174
+ "0.6": 770,
175
+ "0.7": 364,
176
+ "0.8": 245,
177
+ "0.9": 202
178
+ },
179
+ "n_dup_clusters_abs": {
180
+ "0.5": 964,
181
+ "0.6": 312,
182
+ "0.7": 115,
183
+ "0.8": 57,
184
+ "0.9": 44
185
+ },
186
+ "largest_dup_abs": {
187
+ "0.5": 134,
188
+ "0.6": 134,
189
+ "0.7": 133,
190
+ "0.8": 133,
191
+ "0.9": 63
192
+ },
193
+ "babel_mu1_abs": 0.41863396763801575
194
+ }
195
+ },
196
+ "per_pair": {
197
+ "seed0\u2194seed1": {
198
+ "dcs_signed": 0.8013520240783691,
199
+ "dcs_abs": 0.8026490211486816,
200
+ "frac_sign_flipped": 0.014241536458333334,
201
+ "alive_source": "stats",
202
+ "n_alive_a": 12288,
203
+ "n_alive_b": 12288,
204
+ "matched_cos_signed": {
205
+ "0.5": 0.7888076901435852,
206
+ "0.6": 0.7969455718994141,
207
+ "0.7": 0.7999643683433533,
208
+ "0.8": 0.8003712296485901,
209
+ "0.9": 0.8005580902099609
210
+ },
211
+ "unique_frac_signed": {
212
+ "0.5": 0.9132080078125,
213
+ "0.6": 0.9678141276041666,
214
+ "0.7": 0.9854736328125,
215
+ "0.8": 0.9895426432291666,
216
+ "0.9": 0.9913330078125
217
+ },
218
+ "dedup_dcs_signed": {
219
+ "0.5": 0.7203454992632032,
220
+ "0.6": 0.771295183415835,
221
+ "0.7": 0.7883437921918812,
222
+ "0.8": 0.7920014621510442,
223
+ "0.9": 0.7936196594964713
224
+ },
225
+ "matched_frac_signed": {
226
+ "0.5": 0.7733561197916666,
227
+ "0.6": 0.7854817708333334,
228
+ "0.7": 0.7486165364583334,
229
+ "0.8": 0.6634928385416666,
230
+ "0.9": 0.5017903645833334
231
+ },
232
+ "matched_cos_abs": {
233
+ "0.5": 0.7907242774963379,
234
+ "0.6": 0.798725962638855,
235
+ "0.7": 0.8015782237052917,
236
+ "0.8": 0.8020060062408447,
237
+ "0.9": 0.8020365834236145
238
+ },
239
+ "unique_frac_abs": {
240
+ "0.5": 0.9056396484375,
241
+ "0.6": 0.9618733723958334,
242
+ "0.7": 0.9808349609375,
243
+ "0.8": 0.9857177734375,
244
+ "0.9": 0.9887288411458334
245
+ },
246
+ "dedup_dcs_abs": {
247
+ "0.5": 0.7161112566827796,
248
+ "0.6": 0.7682732353035439,
249
+ "0.7": 0.7862159457363305,
250
+ "0.8": 0.7905515747552272,
251
+ "0.9": 0.7929967016849939
252
+ },
253
+ "matched_frac_abs": {
254
+ "0.5": 0.7698567708333334,
255
+ "0.6": 0.783935546875,
256
+ "0.7": 0.746826171875,
257
+ "0.8": 0.662841796875,
258
+ "0.9": 0.5011393229166666
259
+ }
260
+ },
261
+ "seed0\u2194seed2": {
262
+ "dcs_signed": 0.8060774207115173,
263
+ "dcs_abs": 0.807426929473877,
264
+ "frac_sign_flipped": 0.014485677083333334,
265
+ "alive_source": "stats",
266
+ "n_alive_a": 12288,
267
+ "n_alive_b": 12288,
268
+ "matched_cos_signed": {
269
+ "0.5": 0.7918343544006348,
270
+ "0.6": 0.8003867268562317,
271
+ "0.7": 0.8041903376579285,
272
+ "0.8": 0.8048174381256104,
273
+ "0.9": 0.8050143718719482
274
+ },
275
+ "unique_frac_signed": {
276
+ "0.5": 0.9120686848958334,
277
+ "0.6": 0.9678548177083334,
278
+ "0.7": 0.9851888020833334,
279
+ "0.8": 0.9890950520833334,
280
+ "0.9": 0.9903971354166666
281
+ },
282
+ "dedup_dcs_signed": {
283
+ "0.5": 0.7222073182735282,
284
+ "0.6": 0.7746581496176077,
285
+ "0.7": 0.7922793154042059,
286
+ "0.8": 0.7960409458804255,
287
+ "0.9": 0.7972839278712247
288
+ },
289
+ "matched_frac_signed": {
290
+ "0.5": 0.7779134114583334,
291
+ "0.6": 0.7908528645833334,
292
+ "0.7": 0.7503255208333334,
293
+ "0.8": 0.6700846354166666,
294
+ "0.9": 0.5111490885416666
295
+ },
296
+ "matched_cos_abs": {
297
+ "0.5": 0.7933586835861206,
298
+ "0.6": 0.802190899848938,
299
+ "0.7": 0.8059757351875305,
300
+ "0.8": 0.806588351726532,
301
+ "0.9": 0.8066371083259583
302
+ },
303
+ "unique_frac_abs": {
304
+ "0.5": 0.9048665364583334,
305
+ "0.6": 0.962158203125,
306
+ "0.7": 0.9803873697916666,
307
+ "0.8": 0.9851888020833334,
308
+ "0.9": 0.9877522786458334
309
+ },
310
+ "dedup_dcs_abs": {
311
+ "0.5": 0.7178837241857158,
312
+ "0.6": 0.771834554761881,
313
+ "0.7": 0.7901684311364079,
314
+ "0.8": 0.7946418120118324,
315
+ "0.9": 0.7967576417892512
316
+ },
317
+ "matched_frac_abs": {
318
+ "0.5": 0.7736002604166666,
319
+ "0.6": 0.7881673177083334,
320
+ "0.7": 0.7491861979166666,
321
+ "0.8": 0.6689453125,
322
+ "0.9": 0.5105794270833334
323
+ }
324
+ },
325
+ "seed1\u2194seed2": {
326
+ "dcs_signed": 0.8028092384338379,
327
+ "dcs_abs": 0.803881824016571,
328
+ "frac_sign_flipped": 0.013427734375,
329
+ "alive_source": "stats",
330
+ "n_alive_a": 12288,
331
+ "n_alive_b": 12288,
332
+ "matched_cos_signed": {
333
+ "0.5": 0.7915000319480896,
334
+ "0.6": 0.7982300519943237,
335
+ "0.7": 0.8018153309822083,
336
+ "0.8": 0.8023896217346191,
337
+ "0.9": 0.8024964332580566
338
+ },
339
+ "unique_frac_signed": {
340
+ "0.5": 0.9126790364583334,
341
+ "0.6": 0.9682210286458334,
342
+ "0.7": 0.9849853515625,
343
+ "0.8": 0.9892171223958334,
344
+ "0.9": 0.9909261067708334
345
+ },
346
+ "dedup_dcs_signed": {
347
+ "0.5": 0.7223854865151225,
348
+ "0.6": 0.7728631220379611,
349
+ "0.7": 0.7897763556757127,
350
+ "0.8": 0.7937375526526012,
351
+ "0.9": 0.795214666305886
352
+ },
353
+ "matched_frac_signed": {
354
+ "0.5": 0.7759602864583334,
355
+ "0.6": 0.7887369791666666,
356
+ "0.7": 0.7491048177083334,
357
+ "0.8": 0.6639811197916666,
358
+ "0.9": 0.50634765625
359
+ },
360
+ "matched_cos_abs": {
361
+ "0.5": 0.7931916117668152,
362
+ "0.6": 0.8001823425292969,
363
+ "0.7": 0.8033261895179749,
364
+ "0.8": 0.803741455078125,
365
+ "0.9": 0.8037148118019104
366
+ },
367
+ "unique_frac_abs": {
368
+ "0.5": 0.9052327473958334,
369
+ "0.6": 0.9624430338541666,
370
+ "0.7": 0.9801839192708334,
371
+ "0.8": 0.9852294921875,
372
+ "0.9": 0.9881184895833334
373
+ },
374
+ "dedup_dcs_abs": {
375
+ "0.5": 0.7180230219310033,
376
+ "0.6": 0.7701299213804305,
377
+ "0.7": 0.7874074128946329,
378
+ "0.8": 0.7918697856366634,
379
+ "0.9": 0.7941654658934567
380
+ },
381
+ "matched_frac_abs": {
382
+ "0.5": 0.7728678385416666,
383
+ "0.6": 0.784912109375,
384
+ "0.7": 0.74755859375,
385
+ "0.8": 0.663818359375,
386
+ "0.9": 0.505859375
387
+ }
388
+ }
389
+ },
390
+ "summary": {
391
+ "n_seeds": 3,
392
+ "n_pairs": 3,
393
+ "report_band": [
394
+ 0.6,
395
+ 0.7,
396
+ 0.8,
397
+ 0.9
398
+ ],
399
+ "n_dead_mean": 0,
400
+ "n_dead_std": 0.0,
401
+ "n_dense_mean": 21,
402
+ "n_dense_std": 1.1547005383792515,
403
+ "l0_mean": 32,
404
+ "l0_std": 0.06474222029066243,
405
+ "recon_r2_mean": 0.9192662761093985,
406
+ "recon_r2_std": null,
407
+ "l1_per_token_mean": 305.4137613819325,
408
+ "l1_per_token_std": 1.034566429027056,
409
+ "babel_mu1_signed_mean": 0.40890246629714966,
410
+ "babel_mu1_signed_std": 0.00023451253954667774,
411
+ "n_dup_signed_per_threshold": {
412
+ "0.5": 1950.0,
413
+ "0.6": 648.6666666666666,
414
+ "0.7": 236.66666666666666,
415
+ "0.8": 140.33333333333334,
416
+ "0.9": 124.66666666666667
417
+ },
418
+ "n_dup_signed_band": 287.5833333333333,
419
+ "n_dup_signed_band_std": 1.2583057392117918,
420
+ "n_dup_clusters_signed_per_threshold": {
421
+ "0.5": 876.6666666666666,
422
+ "0.6": 255.0,
423
+ "0.7": 55.0,
424
+ "0.8": 8.666666666666666,
425
+ "0.9": 12.666666666666666
426
+ },
427
+ "n_dup_clusters_signed_band": 82.83333333333334,
428
+ "n_dup_clusters_signed_band_std": 4.645786621588784,
429
+ "largest_dup_signed_per_threshold": {
430
+ "0.5": 128.66666666666666,
431
+ "0.6": 128.33333333333334,
432
+ "0.7": 113.33333333333333,
433
+ "0.8": 92.0,
434
+ "0.9": 45.333333333333336
435
+ },
436
+ "largest_dup_signed_band": 94.75,
437
+ "largest_dup_signed_band_std": 24.724734578959588,
438
+ "dcs_signed_mean": 0.8034128944079081,
439
+ "dcs_signed_std": 0.002419843742148464,
440
+ "dedup_dcs_signed_per_threshold": {
441
+ "0.5": 0.721646101350618,
442
+ "0.6": 0.7729388183571345,
443
+ "0.7": 0.7901331544239333,
444
+ "0.8": 0.7939266535613569,
445
+ "0.9": 0.7953727512245273
446
+ },
447
+ "dedup_dcs_signed_band": 0.7880928443917381,
448
+ "dedup_dcs_signed_band_std": 0.001882862504375287,
449
+ "matched_cos_signed_per_threshold": {
450
+ "0.5": 0.7907140254974365,
451
+ "0.6": 0.7985207835833231,
452
+ "0.7": 0.80199001232783,
453
+ "0.8": 0.8025260965029398,
454
+ "0.9": 0.8026896317799886
455
+ },
456
+ "matched_cos_signed_band": 0.8014316310485204,
457
+ "matched_cos_signed_band_std": 0.0020783429698631793,
458
+ "unique_frac_signed_per_threshold": {
459
+ "0.5": 0.9126519097222223,
460
+ "0.6": 0.9679633246527778,
461
+ "0.7": 0.9852159288194445,
462
+ "0.8": 0.9892849392361112,
463
+ "0.9": 0.9908854166666666
464
+ },
465
+ "unique_frac_signed_band": 0.98333740234375,
466
+ "unique_frac_signed_band_std": 0.00020345052083331483,
467
+ "matched_frac_signed_per_threshold": {
468
+ "0.5": 0.7757432725694445,
469
+ "0.6": 0.7883572048611112,
470
+ "0.7": 0.7493489583333334,
471
+ "0.8": 0.6658528645833334,
472
+ "0.9": 0.5064290364583334
473
+ },
474
+ "matched_frac_signed_band": 0.6774970160590279,
475
+ "matched_frac_signed_band_std": 0.0029055935077566897,
476
+ "babel_mu1_abs_mean": 0.41849836707115173,
477
+ "babel_mu1_abs_std": 0.00012127432870639753,
478
+ "n_dup_abs_per_threshold": {
479
+ "0.5": 2124.3333333333335,
480
+ "0.6": 789.0,
481
+ "0.7": 352.0,
482
+ "0.8": 235.66666666666666,
483
+ "0.9": 190.33333333333334
484
+ },
485
+ "n_dup_abs_band": 391.75,
486
+ "n_dup_abs_band_std": 3.774917217635375,
487
+ "n_dup_clusters_abs_per_threshold": {
488
+ "0.5": 960.0,
489
+ "0.6": 324.0,
490
+ "0.7": 112.0,
491
+ "0.8": 56.0,
492
+ "0.9": 45.333333333333336
493
+ },
494
+ "n_dup_clusters_abs_band": 134.33333333333334,
495
+ "n_dup_clusters_abs_band_std": 4.259792639710686,
496
+ "largest_dup_abs_per_threshold": {
497
+ "0.5": 128.66666666666666,
498
+ "0.6": 128.33333333333334,
499
+ "0.7": 113.33333333333333,
500
+ "0.8": 92.0,
501
+ "0.9": 45.333333333333336
502
+ },
503
+ "largest_dup_abs_band": 94.75,
504
+ "largest_dup_abs_band_std": 24.724734578959588,
505
+ "dcs_abs_mean": 0.8046525915463766,
506
+ "dcs_abs_std": 0.0024804563954531683,
507
+ "dedup_dcs_abs_per_threshold": {
508
+ "0.5": 0.7173393342664997,
509
+ "0.6": 0.7700792371486185,
510
+ "0.7": 0.7879305965891238,
511
+ "0.8": 0.792354390801241,
512
+ "0.9": 0.7946399364559006
513
+ },
514
+ "dedup_dcs_abs_band": 0.786251040248721,
515
+ "dedup_dcs_abs_band_std": 0.0019454711076706708,
516
+ "matched_cos_abs_per_threshold": {
517
+ "0.5": 0.7924248576164246,
518
+ "0.6": 0.8003664016723633,
519
+ "0.7": 0.8036267161369324,
520
+ "0.8": 0.8041119376818339,
521
+ "0.9": 0.8041295011838278
522
+ },
523
+ "matched_cos_abs_band": 0.8030586391687393,
524
+ "matched_cos_abs_band_std": 0.0021483269535711945,
525
+ "unique_frac_abs_per_threshold": {
526
+ "0.5": 0.905246310763889,
527
+ "0.6": 0.962158203125,
528
+ "0.7": 0.98046875,
529
+ "0.8": 0.9853786892361112,
530
+ "0.9": 0.9881998697916666
531
+ },
532
+ "unique_frac_abs_band": 0.9790513780381944,
533
+ "unique_frac_abs_band_std": 0.00021442887035812565,
534
+ "matched_frac_abs_per_threshold": {
535
+ "0.5": 0.7721082899305555,
536
+ "0.6": 0.7856716579861112,
537
+ "0.7": 0.7478569878472222,
538
+ "0.8": 0.6652018229166666,
539
+ "0.9": 0.505859375
540
+ },
541
+ "matched_frac_abs_band": 0.6761474609375,
542
+ "matched_frac_abs_band_std": 0.002816963304139469,
543
+ "alive_source": "stats"
544
+ }
545
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 0,
8
+ "top_k": 32,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c6c2199c9edf2af46210a92a87cc0603b21a08c29b150f42e1fe73dbd376c48
3
+ size 75550084
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2714be279f3b7a7edb76b1986f25f063d27243ad7ce03de6156cad09182cbce
3
+ size 99648
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a10b442de0cdce6f5c0c62cda2d5f28e61d69107db1738b5d27495aebed49c7
3
+ size 50471
lm/gpt2-small_layer8/batchtopk/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": 31.96863642578125,
5
+ "l0_per_token_0.001": 31.96863642578125,
6
+ "l0_per_token_0.01": 31.96863642578125,
7
+ "l0_per_token_0.05": 31.96863642578125,
8
+ "l0_per_token_0.1": 31.96863642578125,
9
+ "l0_per_token_0.5": 31.96863642578125,
10
+ "l0_per_token_1.0": 31.96863642578125,
11
+ "l1_per_token": 304.34399717607556,
12
+ "r2_score": 0.9192533434187236,
13
+ "l2_loss": 1.8937056209024563
14
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b8064e976734d923553dc88762712f13d7102d0158b12dee30854a7fe18a887c
3
+ size 50382
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98f7dfb8b1c6e7b94e18ebf154a86be17a59bcefc099499faf1d4c8bbf808277
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d48a6d8329da1058c73a7b0d533ba639f0e04862d2bd5217fb55ec4d605dbc05
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 1,
8
+ "top_k": 32,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f74337b486fc1617db11683aadad32dba731d172d1f30a1bd3637b28051a396b
3
+ size 75550084
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3d7831b6dea24ba39fd2f653f5f34031076d45bb738ef459882af4b2c7b2385
3
+ size 99648
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08bac7104ddae5fd076e1471370107d34af9bdeb2d4d88f3716bf71a8d35748b
3
+ size 50471
lm/gpt2-small_layer8/batchtopk/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": 31.867031298828124,
5
+ "l0_per_token_0.001": 31.867031298828124,
6
+ "l0_per_token_0.01": 31.867031298828124,
7
+ "l0_per_token_0.05": 31.867031298828124,
8
+ "l0_per_token_0.1": 31.867031298828124,
9
+ "l0_per_token_0.5": 31.867031298828124,
10
+ "l0_per_token_1.0": 31.867031298828124,
11
+ "l1_per_token": 306.4091118377043,
12
+ "r2_score": 0.9193055911896232,
13
+ "l2_loss": 1.892480283512387
14
+ }
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9b26d1ce015f466b7e7b329191c8a45245802592d3642d22a21b2e784701d61b
3
+ size 50382
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:350318a196a0217fa49aea67182701743561dcb62cc32a91ab968e4bc0cdd72d
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/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/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8931c98b91e66273ba09c5502ff526cb51526fa467619142f247d773eba40381
3
+ size 50387
lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed2/config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d_in": 768,
3
+ "d_sae_factor": 16,
4
+ "architecture": "batchtopk",
5
+ "activation_name": "relu",
6
+ "device": "cpu",
7
+ "seed": 2,
8
+ "top_k": 32,
9
+ "top_k_inference": 0,
10
+ "class_type": "BatchTopKSAE",
11
+ "module_name": "consdl.models.batchtopk"
12
+ }