chanind commited on
Commit
29472a2
·
verified ·
1 Parent(s): 8730267

Upload scr/gemma-2-2b-res-matryoshka-dc_blocks.15.hook_resid_post_eval_results.json with huggingface_hub

Browse files
scr/gemma-2-2b-res-matryoshka-dc_blocks.15.hook_resid_post_eval_results.json ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "eval_type_id": "scr",
3
+ "eval_config": {
4
+ "random_seed": 42,
5
+ "dataset_names": [
6
+ "LabHC/bias_in_bios_class_set1",
7
+ "canrager/amazon_reviews_mcauley_1and5"
8
+ ],
9
+ "perform_scr": true,
10
+ "early_stopping_patience": 20,
11
+ "train_set_size": 4000,
12
+ "test_set_size": 1000,
13
+ "context_length": 128,
14
+ "probe_train_batch_size": 16,
15
+ "probe_test_batch_size": 500,
16
+ "probe_epochs": 20,
17
+ "probe_lr": 0.001,
18
+ "probe_l1_penalty": 0.001,
19
+ "sae_batch_size": 125,
20
+ "llm_batch_size": 32,
21
+ "llm_dtype": "bfloat16",
22
+ "lower_vram_usage": false,
23
+ "model_name": "gemma-2-2b",
24
+ "n_values": [
25
+ 2,
26
+ 5,
27
+ 10,
28
+ 20,
29
+ 50,
30
+ 100,
31
+ 500
32
+ ],
33
+ "column1_vals_lookup": {
34
+ "LabHC/bias_in_bios_class_set1": [
35
+ [
36
+ "professor",
37
+ "nurse"
38
+ ],
39
+ [
40
+ "architect",
41
+ "journalist"
42
+ ],
43
+ [
44
+ "surgeon",
45
+ "psychologist"
46
+ ],
47
+ [
48
+ "attorney",
49
+ "teacher"
50
+ ]
51
+ ],
52
+ "canrager/amazon_reviews_mcauley_1and5": [
53
+ [
54
+ "Books",
55
+ "CDs_and_Vinyl"
56
+ ],
57
+ [
58
+ "Software",
59
+ "Electronics"
60
+ ],
61
+ [
62
+ "Pet_Supplies",
63
+ "Office_Products"
64
+ ],
65
+ [
66
+ "Industrial_and_Scientific",
67
+ "Toys_and_Games"
68
+ ]
69
+ ]
70
+ }
71
+ },
72
+ "eval_id": "570e36e0-86e6-40dc-b65a-13d96e4b9429",
73
+ "datetime_epoch_millis": 1745749845056,
74
+ "eval_result_metrics": {
75
+ "scr_metrics": {
76
+ "scr_dir1_threshold_2": 0.18415930821903284,
77
+ "scr_metric_threshold_2": 0.11918215737722652,
78
+ "scr_dir2_threshold_2": 0.1615722725617302,
79
+ "scr_dir1_threshold_5": 0.11862516636052428,
80
+ "scr_metric_threshold_5": 0.2130679710749008,
81
+ "scr_dir2_threshold_5": 0.23374973016768538,
82
+ "scr_dir1_threshold_10": 0.12342375834219946,
83
+ "scr_metric_threshold_10": 0.28236007750355363,
84
+ "scr_dir2_threshold_10": 0.2875626940409158,
85
+ "scr_dir1_threshold_20": -0.21917514512721875,
86
+ "scr_metric_threshold_20": 0.31318900805649014,
87
+ "scr_dir2_threshold_20": 0.2995528593320341,
88
+ "scr_dir1_threshold_50": -0.3542758029514693,
89
+ "scr_metric_threshold_50": 0.19915711871286293,
90
+ "scr_dir2_threshold_50": 0.20357748588066388,
91
+ "scr_dir1_threshold_100": -0.4199684902469859,
92
+ "scr_metric_threshold_100": 0.06774996595088081,
93
+ "scr_dir2_threshold_100": 0.14888769321296386,
94
+ "scr_dir1_threshold_500": -1.2285344986691178,
95
+ "scr_metric_threshold_500": 0.0007064118882371521,
96
+ "scr_dir2_threshold_500": 0.046600076724056735
97
+ }
98
+ },
99
+ "eval_result_details": [
100
+ {
101
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results",
102
+ "scr_dir1_threshold_2": 0.40229909469463926,
103
+ "scr_metric_threshold_2": 0.07894729412163301,
104
+ "scr_dir2_threshold_2": 0.07894729412163301,
105
+ "scr_dir1_threshold_5": -0.7356315145469569,
106
+ "scr_metric_threshold_5": 0.12631573333634494,
107
+ "scr_dir2_threshold_5": 0.12631573333634494,
108
+ "scr_dir1_threshold_10": -0.6091950085570335,
109
+ "scr_metric_threshold_10": 0.22368415686562385,
110
+ "scr_dir2_threshold_10": 0.22368415686562385,
111
+ "scr_dir1_threshold_20": -1.1494247124932688,
112
+ "scr_metric_threshold_20": 0.1368420705900954,
113
+ "scr_dir2_threshold_20": 0.1368420705900954,
114
+ "scr_dir1_threshold_50": -1.172413604107376,
115
+ "scr_metric_threshold_50": 0.3815789019632202,
116
+ "scr_dir2_threshold_50": 0.3815789019632202,
117
+ "scr_dir1_threshold_100": 0.0804600929832326,
118
+ "scr_metric_threshold_100": 0.4605263529391835,
119
+ "scr_dir2_threshold_100": 0.4605263529391835,
120
+ "scr_dir1_threshold_500": -3.0804594078724707,
121
+ "scr_metric_threshold_500": 0.2921052705878367,
122
+ "scr_dir2_threshold_500": 0.2921052705878367
123
+ },
124
+ {
125
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results",
126
+ "scr_dir1_threshold_2": 0.2706767591210781,
127
+ "scr_metric_threshold_2": 0.05828219400691707,
128
+ "scr_dir2_threshold_2": 0.05828219400691707,
129
+ "scr_dir1_threshold_5": -0.03759394452735315,
130
+ "scr_metric_threshold_5": 0.1809814728272277,
131
+ "scr_dir2_threshold_5": 0.1809814728272277,
132
+ "scr_dir1_threshold_10": 0.29323285694438844,
133
+ "scr_metric_threshold_10": 0.04907967496085474,
134
+ "scr_dir2_threshold_10": 0.04907967496085474,
135
+ "scr_dir1_threshold_20": -0.052631791231395955,
136
+ "scr_metric_threshold_20": 0.13803668867284946,
137
+ "scr_dir2_threshold_20": 0.13803668867284946,
138
+ "scr_dir1_threshold_50": -1.1879697226367658,
139
+ "scr_metric_threshold_50": -0.16871169121427468,
140
+ "scr_dir2_threshold_50": -0.16871169121427468,
141
+ "scr_dir1_threshold_100": -1.3759398934287008,
142
+ "scr_metric_threshold_100": -0.23312895886401583,
143
+ "scr_dir2_threshold_100": -0.23312895886401583,
144
+ "scr_dir1_threshold_500": -1.8721803197138978,
145
+ "scr_metric_threshold_500": -0.4018406500782905,
146
+ "scr_dir2_threshold_500": -0.4018406500782905
147
+ },
148
+ {
149
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results",
150
+ "scr_dir1_threshold_2": 0.19512157891217186,
151
+ "scr_metric_threshold_2": 0.015831087674656154,
152
+ "scr_dir2_threshold_2": 0.015831087674656154,
153
+ "scr_dir1_threshold_5": -0.085366190508035,
154
+ "scr_metric_threshold_5": 0.09498684058431422,
155
+ "scr_dir2_threshold_5": 0.09498684058431422,
156
+ "scr_dir1_threshold_10": 0.0,
157
+ "scr_metric_threshold_10": 0.21899736811686285,
158
+ "scr_dir2_threshold_10": 0.21899736811686285,
159
+ "scr_dir1_threshold_20": -1.7682920801435162,
160
+ "scr_metric_threshold_20": 0.16622681375436124,
161
+ "scr_dir2_threshold_20": 0.16622681375436124,
162
+ "scr_dir1_threshold_50": -1.4512194235505171,
163
+ "scr_metric_threshold_50": 0.01846957607573407,
164
+ "scr_dir2_threshold_50": 0.01846957607573407,
165
+ "scr_dir1_threshold_100": -1.7804877694202068,
166
+ "scr_metric_threshold_100": -0.0897098637821584,
167
+ "scr_dir2_threshold_100": -0.0897098637821584,
168
+ "scr_dir1_threshold_500": -3.9878043107233094,
169
+ "scr_metric_threshold_500": -0.10817943985789247,
170
+ "scr_dir2_threshold_500": -0.10817943985789247
171
+ },
172
+ {
173
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results",
174
+ "scr_dir1_threshold_2": 0.09929078913605799,
175
+ "scr_metric_threshold_2": -0.040123491993293375,
176
+ "scr_dir2_threshold_2": -0.040123491993293375,
177
+ "scr_dir1_threshold_5": 0.5177302715985503,
178
+ "scr_metric_threshold_5": -0.18827175597122642,
179
+ "scr_dir2_threshold_5": -0.18827175597122642,
180
+ "scr_dir1_threshold_10": -0.4326245451974499,
181
+ "scr_metric_threshold_10": -0.21913592397685133,
182
+ "scr_dir2_threshold_10": -0.21913592397685133,
183
+ "scr_dir1_threshold_20": -0.1631206123469566,
184
+ "scr_metric_threshold_20": -0.3024692879710101,
185
+ "scr_dir2_threshold_20": -0.3024692879710101,
186
+ "scr_dir1_threshold_50": -0.7872346271771024,
187
+ "scr_metric_threshold_50": -0.27469149997295716,
188
+ "scr_dir2_threshold_50": -0.27469149997295716,
189
+ "scr_dir1_threshold_100": -1.2907806814968124,
190
+ "scr_metric_threshold_100": -0.2962963439909136,
191
+ "scr_dir2_threshold_100": -0.2962963439909136,
192
+ "scr_dir1_threshold_500": -1.971631988170378,
193
+ "scr_metric_threshold_500": -0.5493828159809618,
194
+ "scr_dir2_threshold_500": -0.5493828159809618
195
+ },
196
+ {
197
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results",
198
+ "scr_dir1_threshold_2": 0.15811964179140975,
199
+ "scr_metric_threshold_2": 0.4927535731109936,
200
+ "scr_dir2_threshold_2": 0.4927535731109936,
201
+ "scr_dir1_threshold_5": 0.4273504796008221,
202
+ "scr_metric_threshold_5": 0.6280192485553794,
203
+ "scr_dir2_threshold_5": 0.6280192485553794,
204
+ "scr_dir1_threshold_10": 0.47435892537422925,
205
+ "scr_metric_threshold_10": 0.7246375058884778,
206
+ "scr_dir2_threshold_10": 0.7246375058884778,
207
+ "scr_dir1_threshold_20": -0.1752136915419236,
208
+ "scr_metric_threshold_20": 0.8115940526663142,
209
+ "scr_dir2_threshold_20": 0.8115940526663142,
210
+ "scr_dir1_threshold_50": 0.05555559800900132,
211
+ "scr_metric_threshold_50": -0.07246369299982378,
212
+ "scr_dir2_threshold_50": -0.07246369299982378,
213
+ "scr_dir1_threshold_100": 0.1324786945259763,
214
+ "scr_metric_threshold_100": -0.17391309355567303,
215
+ "scr_dir2_threshold_100": -0.17391309355567303,
216
+ "scr_dir1_threshold_500": 0.09829059502494862,
217
+ "scr_metric_threshold_500": -0.21256051166696038,
218
+ "scr_dir2_threshold_500": -0.21256051166696038
219
+ },
220
+ {
221
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results",
222
+ "scr_dir1_threshold_2": 0.10970449172576832,
223
+ "scr_metric_threshold_2": 0.10970449172576832,
224
+ "scr_dir2_threshold_2": 0.18137255474862032,
225
+ "scr_dir1_threshold_5": 0.30801669936119913,
226
+ "scr_metric_threshold_5": 0.30801669936119913,
227
+ "scr_dir2_threshold_5": 0.27941167877069506,
228
+ "scr_dir1_threshold_10": 0.4219407977465922,
229
+ "scr_metric_threshold_10": 0.4219407977465922,
230
+ "scr_dir2_threshold_10": 0.294117664245861,
231
+ "scr_dir1_threshold_20": 0.5780589507570042,
232
+ "scr_metric_threshold_20": 0.5780589507570042,
233
+ "scr_dir2_threshold_20": 0.31862744525137965,
234
+ "scr_dir1_threshold_50": 0.6371306775313113,
235
+ "scr_metric_threshold_50": 0.6371306775313113,
236
+ "scr_dir2_threshold_50": 0.5098040877099895,
237
+ "scr_dir1_threshold_100": 0.7046413661284644,
238
+ "scr_metric_threshold_100": 0.7046413661284644,
239
+ "scr_dir2_threshold_100": 0.38725489050275935,
240
+ "scr_dir1_threshold_500": 0.5021095518334088,
241
+ "scr_metric_threshold_500": 0.5021095518334088,
242
+ "scr_dir2_threshold_500": 0.4166668614530911
243
+ },
244
+ {
245
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results",
246
+ "scr_dir1_threshold_2": 0.16883131292515077,
247
+ "scr_metric_threshold_2": 0.16883131292515077,
248
+ "scr_dir2_threshold_2": 0.3899081678324661,
249
+ "scr_dir1_threshold_5": 0.35064947128617274,
250
+ "scr_metric_threshold_5": 0.35064947128617274,
251
+ "scr_dir2_threshold_5": 0.49999986329209584,
252
+ "scr_dir1_threshold_10": 0.528138681168571,
253
+ "scr_metric_threshold_10": 0.528138681168571,
254
+ "scr_dir2_threshold_10": 0.5871558002566281,
255
+ "scr_dir1_threshold_20": 0.6233765798133265,
256
+ "scr_metric_threshold_20": 0.6233765798133265,
257
+ "scr_dir2_threshold_20": 0.6651375431894446,
258
+ "scr_dir1_threshold_50": 0.6450215802352035,
259
+ "scr_metric_threshold_50": 0.6450215802352035,
260
+ "scr_dir2_threshold_50": 0.7201835276271636,
261
+ "scr_dir1_threshold_100": 0.35064947128617274,
262
+ "scr_metric_threshold_100": 0.35064947128617274,
263
+ "scr_dir2_threshold_100": 0.7201835276271636,
264
+ "scr_dir1_threshold_500": 0.47186157686018737,
265
+ "scr_metric_threshold_500": 0.47186157686018737,
266
+ "scr_dir2_threshold_500": 0.6238531232151074
267
+ },
268
+ {
269
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results",
270
+ "scr_dir1_threshold_2": 0.06923079744598673,
271
+ "scr_metric_threshold_2": 0.06923079744598673,
272
+ "scr_dir2_threshold_2": 0.11560680099184864,
273
+ "scr_dir1_threshold_5": 0.2038460586197948,
274
+ "scr_metric_threshold_5": 0.2038460586197948,
275
+ "scr_dir2_threshold_5": 0.24855475994665213,
276
+ "scr_dir1_threshold_10": 0.3115383592582981,
277
+ "scr_metric_threshold_10": 0.3115383592582981,
278
+ "scr_dir2_threshold_10": 0.421965305969869,
279
+ "scr_dir1_threshold_20": 0.3538461961689801,
280
+ "scr_metric_threshold_20": 0.3538461961689801,
281
+ "scr_dir2_threshold_20": 0.46242754850283846,
282
+ "scr_dir1_threshold_50": 0.42692309808449,
283
+ "scr_metric_threshold_50": 0.42692309808449,
284
+ "scr_dir2_threshold_50": 0.5144506778562591,
285
+ "scr_dir1_threshold_100": -0.18076920255401327,
286
+ "scr_metric_threshold_100": -0.18076920255401327,
287
+ "scr_dir2_threshold_100": 0.41618503482736535,
288
+ "scr_dir1_threshold_500": 0.011538313408569681,
289
+ "scr_metric_threshold_500": 0.011538313408569681,
290
+ "scr_dir2_threshold_500": 0.312138776120524
291
+ }
292
+ ],
293
+ "sae_bench_commit_hash": "Unknown",
294
+ "sae_lens_id": "blocks.15.hook_resid_post",
295
+ "sae_lens_release_id": "gemma-2-2b-res-matryoshka-dc",
296
+ "sae_lens_version": "5.9.1",
297
+ "sae_cfg_dict": {
298
+ "architecture": "jumprelu",
299
+ "d_in": 2304,
300
+ "d_sae": 32768,
301
+ "activation_fn_str": "relu",
302
+ "apply_b_dec_to_input": true,
303
+ "finetuning_scaling_factor": false,
304
+ "context_size": 1024,
305
+ "model_name": "gemma-2-2b",
306
+ "hook_name": "blocks.15.hook_resid_post",
307
+ "hook_layer": 15,
308
+ "hook_head_index": null,
309
+ "prepend_bos": true,
310
+ "dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B",
311
+ "dataset_trust_remote_code": true,
312
+ "normalize_activations": "none",
313
+ "dtype": "torch.bfloat16",
314
+ "device": "cuda",
315
+ "sae_lens_training_version": "5.5.1",
316
+ "activation_fn_kwargs": {
317
+ "k": 40
318
+ },
319
+ "neuronpedia_id": null,
320
+ "model_from_pretrained_kwargs": {
321
+ "center_writing_weights": false
322
+ },
323
+ "seqpos_slice": [
324
+ null
325
+ ]
326
+ },
327
+ "eval_result_unstructured": null
328
+ }