chanind commited on
Commit
8730267
·
verified ·
1 Parent(s): 8de4a09

Upload scr/gemma-2-2b-res-snap-matryoshka-dc_blocks.14.hook_resid_post_eval_results.json with huggingface_hub

Browse files
scr/gemma-2-2b-res-snap-matryoshka-dc_blocks.14.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": "404efc13-d102-4864-9ae1-43c4ff9dca0f",
73
+ "datetime_epoch_millis": 1745608503156,
74
+ "eval_result_metrics": {
75
+ "scr_metrics": {
76
+ "scr_dir1_threshold_2": 0.2563807450169605,
77
+ "scr_metric_threshold_2": 0.15203619788306122,
78
+ "scr_dir2_threshold_2": 0.2413469166888173,
79
+ "scr_dir1_threshold_5": 0.04394060360136729,
80
+ "scr_metric_threshold_5": 0.22113237686460954,
81
+ "scr_dir2_threshold_5": 0.291148238989286,
82
+ "scr_dir1_threshold_10": 0.10644533197310384,
83
+ "scr_metric_threshold_10": 0.2915860223172376,
84
+ "scr_dir2_threshold_10": 0.3475553066166193,
85
+ "scr_dir1_threshold_20": 0.11473714651497255,
86
+ "scr_metric_threshold_20": 0.34665754510806035,
87
+ "scr_dir2_threshold_20": 0.3696919182805279,
88
+ "scr_dir1_threshold_50": -0.23295781107641975,
89
+ "scr_metric_threshold_50": 0.30968446924921256,
90
+ "scr_dir2_threshold_50": 0.30178125427389885,
91
+ "scr_dir1_threshold_100": -0.6025557558117175,
92
+ "scr_metric_threshold_100": 0.31622342493097577,
93
+ "scr_dir2_threshold_100": 0.34326894701612665,
94
+ "scr_dir1_threshold_500": -1.080635788361094,
95
+ "scr_metric_threshold_500": 0.03231900125560101,
96
+ "scr_dir2_threshold_500": 0.21430634559026918
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": 0.47368454074606214,
103
+ "scr_metric_threshold_2": 0.03466653823852539,
104
+ "scr_dir2_threshold_2": 0.03466653823852539,
105
+ "scr_dir1_threshold_5": 0.17894700517933163,
106
+ "scr_metric_threshold_5": 0.15199995040893555,
107
+ "scr_dir2_threshold_5": 0.15199995040893555,
108
+ "scr_dir1_threshold_10": 0.4526315459253938,
109
+ "scr_metric_threshold_10": 0.23199987411499023,
110
+ "scr_dir2_threshold_10": 0.23199987411499023,
111
+ "scr_dir1_threshold_20": 0.4210523674031503,
112
+ "scr_metric_threshold_20": 0.3066665331522624,
113
+ "scr_dir2_threshold_20": 0.3066665331522624,
114
+ "scr_dir1_threshold_50": 0.12631545925393783,
115
+ "scr_metric_threshold_50": 0.41066662470499676,
116
+ "scr_dir2_threshold_50": 0.41066662470499676,
117
+ "scr_dir1_threshold_100": -1.0210529948206684,
118
+ "scr_metric_threshold_100": 0.48799991607666016,
119
+ "scr_dir2_threshold_100": 0.48799991607666016,
120
+ "scr_dir1_threshold_500": -1.8315798059397617,
121
+ "scr_metric_threshold_500": 0.1706665356953939,
122
+ "scr_dir2_threshold_500": 0.1706665356953939
123
+ },
124
+ {
125
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results",
126
+ "scr_dir1_threshold_2": 0.41228075679101034,
127
+ "scr_metric_threshold_2": 0.14369507873668225,
128
+ "scr_dir2_threshold_2": 0.14369507873668225,
129
+ "scr_dir1_threshold_5": -0.3947371172883851,
130
+ "scr_metric_threshold_5": 0.1730205622028617,
131
+ "scr_dir2_threshold_5": 0.1730205622028617,
132
+ "scr_dir1_threshold_10": -0.956140901243437,
133
+ "scr_metric_threshold_10": 0.30791792598222245,
134
+ "scr_dir2_threshold_10": 0.30791792598222245,
135
+ "scr_dir1_threshold_20": -0.9122812796389421,
136
+ "scr_metric_threshold_20": 0.3049854125943383,
137
+ "scr_dir2_threshold_20": 0.3049854125943383,
138
+ "scr_dir1_threshold_50": -1.6754395321138429,
139
+ "scr_metric_threshold_50": -0.052785940156590654,
140
+ "scr_dir2_threshold_50": -0.052785940156590654,
141
+ "scr_dir1_threshold_100": -0.5087723425992444,
142
+ "scr_metric_threshold_100": -0.06451616850179627,
143
+ "scr_dir2_threshold_100": -0.06451616850179627,
144
+ "scr_dir1_threshold_500": -1.9210536222381864,
145
+ "scr_metric_threshold_500": -0.2492667842561944,
146
+ "scr_dir2_threshold_500": -0.2492667842561944
147
+ },
148
+ {
149
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results",
150
+ "scr_dir1_threshold_2": 0.5,
151
+ "scr_metric_threshold_2": 0.22222227356127566,
152
+ "scr_dir2_threshold_2": 0.22222227356127566,
153
+ "scr_dir1_threshold_5": 0.5263162435258355,
154
+ "scr_metric_threshold_5": 0.20671829013845064,
155
+ "scr_dir2_threshold_5": 0.20671829013845064,
156
+ "scr_dir1_threshold_10": 0.5263162435258355,
157
+ "scr_metric_threshold_10": 0.29715762870867546,
158
+ "scr_dir2_threshold_10": 0.29715762870867546,
159
+ "scr_dir1_threshold_20": 0.32894716203371116,
160
+ "scr_metric_threshold_20": 0.22997411125552786,
161
+ "scr_dir2_threshold_20": 0.22997411125552786,
162
+ "scr_dir1_threshold_50": -2.552632487051671,
163
+ "scr_metric_threshold_50": 0.08527129275689639,
164
+ "scr_dir2_threshold_50": 0.08527129275689639,
165
+ "scr_dir1_threshold_100": -3.052632487051671,
166
+ "scr_metric_threshold_100": -0.05167953403032259,
167
+ "scr_dir2_threshold_100": -0.05167953403032259,
168
+ "scr_dir1_threshold_500": -4.197369865764022,
169
+ "scr_metric_threshold_500": 0.0,
170
+ "scr_dir2_threshold_500": 0.0
171
+ },
172
+ {
173
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results",
174
+ "scr_dir1_threshold_2": 0.3308268014717416,
175
+ "scr_metric_threshold_2": 0.08108093111403586,
176
+ "scr_dir2_threshold_2": 0.08108093111403586,
177
+ "scr_dir1_threshold_5": -0.8496242218905874,
178
+ "scr_metric_threshold_5": 0.11411407541294115,
179
+ "scr_dir2_threshold_5": 0.11411407541294115,
180
+ "scr_dir1_threshold_10": -0.5112782729892398,
181
+ "scr_metric_threshold_10": -0.02702703670232027,
182
+ "scr_dir2_threshold_10": -0.02702703670232027,
183
+ "scr_dir1_threshold_20": -0.548872217516593,
184
+ "scr_metric_threshold_20": 0.021020929105735237,
185
+ "scr_dir2_threshold_20": 0.021020929105735237,
186
+ "scr_dir1_threshold_50": 0.2857141576699516,
187
+ "scr_metric_threshold_50": -0.024024072400490225,
188
+ "scr_dir2_threshold_50": -0.024024072400490225,
189
+ "scr_dir1_threshold_100": -1.8120302773632342,
190
+ "scr_metric_threshold_100": -0.0330331442989053,
191
+ "scr_dir2_threshold_100": -0.0330331442989053,
192
+ "scr_dir1_threshold_500": -1.0751878890547062,
193
+ "scr_metric_threshold_500": -0.3813814781343138,
194
+ "scr_dir2_threshold_500": -0.3813814781343138
195
+ },
196
+ {
197
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results",
198
+ "scr_dir1_threshold_2": 0.11764715401326675,
199
+ "scr_metric_threshold_2": 0.5180180543003676,
200
+ "scr_dir2_threshold_2": 0.5180180543003676,
201
+ "scr_dir1_threshold_5": 0.36651594693838474,
202
+ "scr_metric_threshold_5": 0.5990991644073285,
203
+ "scr_dir2_threshold_5": 0.5990991644073285,
204
+ "scr_dir1_threshold_10": 0.4615384200454991,
205
+ "scr_metric_threshold_10": 0.644144165913554,
206
+ "scr_dir2_threshold_10": 0.644144165913554,
207
+ "scr_dir1_threshold_20": 0.49321275755070776,
208
+ "scr_metric_threshold_20": 0.7747749924688727,
209
+ "scr_dir2_threshold_20": 0.7747749924688727,
210
+ "scr_dir1_threshold_50": 0.47511317464833935,
211
+ "scr_metric_threshold_50": 0.5810811101069608,
212
+ "scr_dir2_threshold_50": 0.5810811101069608,
213
+ "scr_dir1_threshold_100": 0.03167433750520866,
214
+ "scr_metric_threshold_100": 0.648648612366299,
215
+ "scr_dir2_threshold_100": 0.648648612366299,
216
+ "scr_dir1_threshold_500": -0.1312216389118512,
217
+ "scr_metric_threshold_500": 0.20720722172014708,
218
+ "scr_dir2_threshold_500": 0.20720722172014708
219
+ },
220
+ {
221
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results",
222
+ "scr_dir1_threshold_2": 0.05957447348163479,
223
+ "scr_metric_threshold_2": 0.05957447348163479,
224
+ "scr_dir2_threshold_2": 0.28301886261973647,
225
+ "scr_dir1_threshold_5": 0.18723395081220856,
226
+ "scr_metric_threshold_5": 0.18723395081220856,
227
+ "scr_dir2_threshold_5": 0.3207547859434262,
228
+ "scr_dir1_threshold_10": 0.29787236740817397,
229
+ "scr_metric_threshold_10": 0.29787236740817397,
230
+ "scr_dir2_threshold_10": 0.33018862619736444,
231
+ "scr_dir1_threshold_20": 0.40425526518365207,
232
+ "scr_metric_threshold_20": 0.40425526518365207,
233
+ "scr_dir2_threshold_20": 0.3018868242815813,
234
+ "scr_dir1_threshold_50": 0.5446807917020173,
235
+ "scr_metric_threshold_50": 0.5446807917020173,
236
+ "scr_dir2_threshold_50": 0.4103773929717131,
237
+ "scr_dir1_threshold_100": 0.5914894696326958,
238
+ "scr_metric_threshold_100": 0.5914894696326958,
239
+ "scr_dir2_threshold_100": 0.6745282939296047,
240
+ "scr_dir1_threshold_500": 0.4723405226694262,
241
+ "scr_metric_threshold_500": 0.4723405226694262,
242
+ "scr_dir2_threshold_500": 0.6839624153375112
243
+ },
244
+ {
245
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results",
246
+ "scr_dir1_threshold_2": 0.09205026242251708,
247
+ "scr_metric_threshold_2": 0.09205026242251708,
248
+ "scr_dir2_threshold_2": 0.5097087561235807,
249
+ "scr_dir1_threshold_5": 0.1924686171599013,
250
+ "scr_metric_threshold_5": 0.1924686171599013,
251
+ "scr_dir2_threshold_5": 0.5873785157693324,
252
+ "scr_dir1_threshold_10": 0.364016754140839,
253
+ "scr_metric_threshold_10": 0.364016754140839,
254
+ "scr_dir2_threshold_10": 0.6941746884572729,
255
+ "scr_dir1_threshold_20": 0.46443510887822326,
256
+ "scr_metric_threshold_20": 0.46443510887822326,
257
+ "scr_dir2_threshold_20": 0.8106794725973473,
258
+ "scr_dir1_threshold_50": 0.6401672920165945,
259
+ "scr_metric_threshold_50": 0.6401672920165945,
260
+ "scr_dir2_threshold_50": 0.7961165554191465,
261
+ "scr_dir1_threshold_100": 0.6945606402386181,
262
+ "scr_metric_threshold_100": 0.6945606402386181,
263
+ "scr_dir2_threshold_100": 0.8640777036127643,
264
+ "scr_dir1_threshold_500": -0.1631800446660706,
265
+ "scr_metric_threshold_500": -0.1631800446660706,
266
+ "scr_dir2_threshold_500": 0.786407654624119
267
+ },
268
+ {
269
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results",
270
+ "scr_dir1_threshold_2": 0.06498197120945136,
271
+ "scr_metric_threshold_2": 0.06498197120945136,
272
+ "scr_dir2_threshold_2": 0.13836483881633452,
273
+ "scr_dir1_threshold_5": 0.14440440437424915,
274
+ "scr_metric_threshold_5": 0.14440440437424915,
275
+ "scr_dir2_threshold_5": 0.1761005676310121,
276
+ "scr_dir1_threshold_10": 0.21660649897176612,
277
+ "scr_metric_threshold_10": 0.21660649897176612,
278
+ "scr_dir2_threshold_10": 0.30188658026119575,
279
+ "scr_dir1_threshold_20": 0.26714800822587104,
280
+ "scr_metric_threshold_20": 0.26714800822587104,
281
+ "scr_dir2_threshold_20": 0.207547070788558,
282
+ "scr_dir1_threshold_50": 0.2924186552633159,
283
+ "scr_metric_threshold_50": 0.2924186552633159,
284
+ "scr_dir2_threshold_50": 0.207547070788558,
285
+ "scr_dir1_threshold_100": 0.2563176079645574,
286
+ "scr_metric_threshold_100": 0.2563176079645574,
287
+ "scr_dir2_threshold_100": 0.22012589697470888,
288
+ "scr_dir1_threshold_500": 0.2021660370164197,
289
+ "scr_metric_threshold_500": 0.2021660370164197,
290
+ "scr_dir2_threshold_500": 0.4968551997354904
291
+ }
292
+ ],
293
+ "sae_bench_commit_hash": "Unknown",
294
+ "sae_lens_id": "blocks.14.hook_resid_post",
295
+ "sae_lens_release_id": "gemma-2-2b-res-snap-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.14.hook_resid_post",
307
+ "hook_layer": 14,
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
+ }