chanind commited on
Commit
7b0f2f6
·
verified ·
1 Parent(s): 72a8421

Upload scr/gemma-2-2b-res-matryoshka-dc_blocks.4.hook_resid_post_eval_results.json with huggingface_hub

Browse files
scr/gemma-2-2b-res-matryoshka-dc_blocks.4.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": 1745748046029,
74
+ "eval_result_metrics": {
75
+ "scr_metrics": {
76
+ "scr_dir1_threshold_2": 0.0667310906271201,
77
+ "scr_metric_threshold_2": 0.09957714807289104,
78
+ "scr_dir2_threshold_2": 0.09957714807289104,
79
+ "scr_dir1_threshold_5": -0.24114705244359289,
80
+ "scr_metric_threshold_5": 0.15609425248446604,
81
+ "scr_dir2_threshold_5": 0.15609425248446604,
82
+ "scr_dir1_threshold_10": -0.31787294353857154,
83
+ "scr_metric_threshold_10": 0.17088989846646446,
84
+ "scr_dir2_threshold_10": 0.17088989846646446,
85
+ "scr_dir1_threshold_20": -0.2539458150198963,
86
+ "scr_metric_threshold_20": 0.24466202181657848,
87
+ "scr_dir2_threshold_20": 0.24466202181657848,
88
+ "scr_dir1_threshold_50": -0.6173371973184318,
89
+ "scr_metric_threshold_50": 0.22971198196061834,
90
+ "scr_dir2_threshold_50": 0.22971198196061834,
91
+ "scr_dir1_threshold_100": -0.7572770388945369,
92
+ "scr_metric_threshold_100": 0.22954331505737816,
93
+ "scr_dir2_threshold_100": 0.22954331505737816,
94
+ "scr_dir1_threshold_500": -3.1647399576480697,
95
+ "scr_metric_threshold_500": -9.420276474522704e-05,
96
+ "scr_dir2_threshold_500": -9.420276474522704e-05
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": 0.03225955308078924,
103
+ "scr_metric_threshold_2": 0.06588234436140777,
104
+ "scr_dir2_threshold_2": 0.06588234436140777,
105
+ "scr_dir1_threshold_5": -1.7096755586490135,
106
+ "scr_metric_threshold_5": 0.0847059113921583,
107
+ "scr_dir2_threshold_5": 0.0847059113921583,
108
+ "scr_dir1_threshold_10": -1.3870954096759431,
109
+ "scr_metric_threshold_10": 0.11294112169206502,
110
+ "scr_dir2_threshold_10": 0.11294112169206502,
111
+ "scr_dir1_threshold_20": -0.7741927420812391,
112
+ "scr_metric_threshold_20": 0.1788234660534728,
113
+ "scr_dir2_threshold_20": 0.1788234660534728,
114
+ "scr_dir1_threshold_50": -1.1612881517571823,
115
+ "scr_metric_threshold_50": 0.14352938305547985,
116
+ "scr_dir2_threshold_50": 0.14352938305547985,
117
+ "scr_dir1_threshold_100": -2.806448449703323,
118
+ "scr_metric_threshold_100": 0.1788234660534728,
119
+ "scr_dir2_threshold_100": 0.1788234660534728,
120
+ "scr_dir1_threshold_500": -8.58063926905521,
121
+ "scr_metric_threshold_500": -0.10823530005748692,
122
+ "scr_dir2_threshold_500": -0.10823530005748692
123
+ },
124
+ {
125
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results",
126
+ "scr_dir1_threshold_2": -0.08333395421533346,
127
+ "scr_metric_threshold_2": 0.04188489844664378,
128
+ "scr_dir2_threshold_2": 0.04188489844664378,
129
+ "scr_dir1_threshold_5": 0.2222219462746666,
130
+ "scr_metric_threshold_5": 0.18062827347669966,
131
+ "scr_dir2_threshold_5": 0.18062827347669966,
132
+ "scr_dir1_threshold_10": -0.47222298107800015,
133
+ "scr_metric_threshold_10": 0.146596832497074,
134
+ "scr_dir2_threshold_10": 0.146596832497074,
135
+ "scr_dir1_threshold_20": 0.05555486568666652,
136
+ "scr_metric_threshold_20": 0.3403140977300774,
137
+ "scr_dir2_threshold_20": 0.3403140977300774,
138
+ "scr_dir1_threshold_50": 0.18055496916699987,
139
+ "scr_metric_threshold_50": 0.18848173094371778,
140
+ "scr_dir2_threshold_50": 0.18848173094371778,
141
+ "scr_dir1_threshold_100": -0.7916675979896669,
142
+ "scr_metric_threshold_100": 0.09424086547185889,
143
+ "scr_dir2_threshold_100": 0.09424086547185889,
144
+ "scr_dir1_threshold_500": -3.2083357133810004,
145
+ "scr_metric_threshold_500": -0.2617799911591651,
146
+ "scr_dir2_threshold_500": -0.2617799911591651
147
+ },
148
+ {
149
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results",
150
+ "scr_dir1_threshold_2": 0.44444518030398356,
151
+ "scr_metric_threshold_2": 0.04433484302398132,
152
+ "scr_dir2_threshold_2": 0.04433484302398132,
153
+ "scr_dir1_threshold_5": -0.962963453535989,
154
+ "scr_metric_threshold_5": 0.08374378812458251,
155
+ "scr_dir2_threshold_5": 0.08374378812458251,
156
+ "scr_dir1_threshold_10": -1.6666666666666667,
157
+ "scr_metric_threshold_10": 0.12561575559160543,
158
+ "scr_dir2_threshold_10": 0.12561575559160543,
159
+ "scr_dir1_threshold_20": -2.2592580328266942,
160
+ "scr_metric_threshold_20": 0.11576351931645514,
161
+ "scr_dir2_threshold_20": 0.11576351931645514,
162
+ "scr_dir1_threshold_50": -5.074073092928022,
163
+ "scr_metric_threshold_50": 0.1403940365995992,
164
+ "scr_dir2_threshold_50": 0.1403940365995992,
165
+ "scr_dir1_threshold_100": -3.3333333333333335,
166
+ "scr_metric_threshold_100": 0.22660084709060346,
167
+ "scr_dir2_threshold_100": 0.22660084709060346,
168
+ "scr_dir1_threshold_500": -10.037034338885393,
169
+ "scr_metric_threshold_500": 0.23152703863291027,
170
+ "scr_dir2_threshold_500": 0.23152703863291027
171
+ },
172
+ {
173
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results",
174
+ "scr_dir1_threshold_2": -0.10891065736834309,
175
+ "scr_metric_threshold_2": 0.057692380119756206,
176
+ "scr_dir2_threshold_2": 0.057692380119756206,
177
+ "scr_dir1_threshold_5": 0.1287129706184542,
178
+ "scr_metric_threshold_5": 0.18406600919254215,
179
+ "scr_dir2_threshold_5": 0.18406600919254215,
180
+ "scr_dir1_threshold_10": 0.19801959163150878,
181
+ "scr_metric_threshold_10": 0.038461532163499464,
182
+ "scr_dir2_threshold_10": 0.038461532163499464,
183
+ "scr_dir1_threshold_20": 0.019802313250111094,
184
+ "scr_metric_threshold_20": 0.29945060568304055,
185
+ "scr_dir2_threshold_20": 0.29945060568304055,
186
+ "scr_dir1_threshold_50": -0.22772276643420858,
187
+ "scr_metric_threshold_50": 0.09890112949157405,
188
+ "scr_dir2_threshold_50": 0.09890112949157405,
189
+ "scr_dir1_threshold_100": 0.08910893426316568,
190
+ "scr_metric_threshold_100": 0.11813197744783079,
191
+ "scr_dir2_threshold_100": 0.11813197744783079,
192
+ "scr_dir1_threshold_500": -0.7128714741053255,
193
+ "scr_metric_threshold_500": -0.32142850709860166,
194
+ "scr_dir2_threshold_500": -0.32142850709860166
195
+ },
196
+ {
197
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results",
198
+ "scr_dir1_threshold_2": -0.027972083353132077,
199
+ "scr_metric_threshold_2": 0.3277311639911775,
200
+ "scr_dir2_threshold_2": 0.3277311639911775,
201
+ "scr_dir1_threshold_5": -0.034964999987495526,
202
+ "scr_metric_threshold_5": 0.4117647058823529,
203
+ "scr_dir2_threshold_5": 0.4117647058823529,
204
+ "scr_dir1_threshold_10": 0.12587416668125523,
205
+ "scr_metric_threshold_10": 0.4663866082875008,
206
+ "scr_dir2_threshold_10": 0.4663866082875008,
207
+ "scr_dir1_threshold_20": -0.034964999987495526,
208
+ "scr_metric_threshold_20": 0.3403360450110281,
209
+ "scr_dir2_threshold_20": 0.3403360450110281,
210
+ "scr_dir1_threshold_50": 0.1048949999624866,
211
+ "scr_metric_threshold_50": 0.4537814768279407,
212
+ "scr_dir2_threshold_50": 0.4537814768279407,
213
+ "scr_dir1_threshold_100": 0.034964999987495526,
214
+ "scr_metric_threshold_100": 0.31512603253161736,
215
+ "scr_dir2_threshold_100": 0.31512603253161736,
216
+ "scr_dir1_threshold_500": -1.5104895833593843,
217
+ "scr_metric_threshold_500": -0.06722703386470796,
218
+ "scr_dir2_threshold_500": -0.06722703386470796
219
+ },
220
+ {
221
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results",
222
+ "scr_dir1_threshold_2": 0.00980379553035279,
223
+ "scr_metric_threshold_2": 0.20599252106867771,
224
+ "scr_dir2_threshold_2": 0.20599252106867771,
225
+ "scr_dir1_threshold_5": 0.00980379553035279,
226
+ "scr_metric_threshold_5": 0.21722855388203682,
227
+ "scr_dir2_threshold_5": 0.21722855388203682,
228
+ "scr_dir1_threshold_10": 0.17647065698344394,
229
+ "scr_metric_threshold_10": 0.2359551264120663,
230
+ "scr_dir2_threshold_10": 0.2359551264120663,
231
+ "scr_dir1_threshold_20": 0.2647059854751659,
232
+ "scr_metric_threshold_20": 0.29962560695717866,
233
+ "scr_dir2_threshold_20": 0.29962560695717866,
234
+ "scr_dir1_threshold_50": 0.4215684670386308,
235
+ "scr_metric_threshold_50": 0.34082402187557276,
236
+ "scr_dir2_threshold_50": 0.34082402187557276,
237
+ "scr_dir1_threshold_100": 0.3333337229061823,
238
+ "scr_metric_threshold_100": 0.4569289501524383,
239
+ "scr_dir2_threshold_100": 0.4569289501524383,
240
+ "scr_dir1_threshold_500": 0.5294119709503318,
241
+ "scr_metric_threshold_500": 0.45318368029410316,
242
+ "scr_dir2_threshold_500": 0.45318368029410316
243
+ },
244
+ {
245
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results",
246
+ "scr_dir1_threshold_2": 0.168141574248678,
247
+ "scr_metric_threshold_2": 0.08996546643413296,
248
+ "scr_dir2_threshold_2": 0.08996546643413296,
249
+ "scr_dir1_threshold_5": 0.2831859480701014,
250
+ "scr_metric_threshold_5": 0.18339109552086616,
251
+ "scr_dir2_threshold_5": 0.18339109552086616,
252
+ "scr_dir1_threshold_10": 0.3539820394814922,
253
+ "scr_metric_threshold_10": 0.3287197831298663,
254
+ "scr_dir2_threshold_10": 0.3287197831298663,
255
+ "scr_dir1_threshold_20": 0.43362810385979716,
256
+ "scr_metric_threshold_20": 0.42906573752180005,
257
+ "scr_dir2_threshold_20": 0.42906573752180005,
258
+ "scr_dir1_threshold_50": 0.5132741682381021,
259
+ "scr_metric_threshold_50": 0.5086505097537112,
260
+ "scr_dir2_threshold_50": 0.5086505097537112,
261
+ "scr_dir1_threshold_100": 0.07079609141139082,
262
+ "scr_metric_threshold_100": 0.5570934056233778,
263
+ "scr_dir2_threshold_100": 0.5570934056233778,
264
+ "scr_dir1_threshold_500": -1.3008853665290836,
265
+ "scr_metric_threshold_500": 0.5432525487685558,
266
+ "scr_dir2_threshold_500": 0.5432525487685558
267
+ },
268
+ {
269
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results",
270
+ "scr_dir1_threshold_2": 0.09941531678996579,
271
+ "scr_metric_threshold_2": -0.03686643286264893,
272
+ "scr_dir2_threshold_2": -0.03686643286264893,
273
+ "scr_dir1_threshold_5": 0.13450293213017933,
274
+ "scr_metric_threshold_5": -0.09677431759551026,
275
+ "scr_dir2_threshold_5": -0.09677431759551026,
276
+ "scr_dir1_threshold_10": 0.12865505433433813,
277
+ "scr_metric_threshold_10": -0.08755757204196167,
278
+ "scr_dir2_threshold_10": -0.08755757204196167,
279
+ "scr_dir1_threshold_20": 0.26315798646451743,
280
+ "scr_metric_threshold_20": -0.0460829037404248,
281
+ "scr_dir2_threshold_20": -0.0460829037404248,
282
+ "scr_dir1_threshold_50": 0.30409382816573854,
283
+ "scr_metric_threshold_50": -0.03686643286264893,
284
+ "scr_dir2_threshold_50": -0.03686643286264893,
285
+ "scr_dir1_threshold_100": 0.3450293213017933,
286
+ "scr_metric_threshold_100": -0.11059902391217406,
287
+ "scr_dir2_threshold_100": -0.11059902391217406,
288
+ "scr_dir1_threshold_500": -0.4970758868194962,
289
+ "scr_metric_threshold_500": -0.4700460576335693,
290
+ "scr_dir2_threshold_500": -0.4700460576335693
291
+ }
292
+ ],
293
+ "sae_bench_commit_hash": "Unknown",
294
+ "sae_lens_id": "blocks.4.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.4.hook_resid_post",
307
+ "hook_layer": 4,
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
+ }