chanind commited on
Commit
cd02f71
·
verified ·
1 Parent(s): 5ffe903

Upload scr/gemma-2-2b-res-matryoshka-dc_blocks.10.hook_resid_post_eval_results.json with huggingface_hub

Browse files
scr/gemma-2-2b-res-matryoshka-dc_blocks.10.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": 1745749035127,
74
+ "eval_result_metrics": {
75
+ "scr_metrics": {
76
+ "scr_dir1_threshold_2": 0.24991608374644733,
77
+ "scr_metric_threshold_2": 0.10917593231900458,
78
+ "scr_dir2_threshold_2": 0.10917593231900458,
79
+ "scr_dir1_threshold_5": -0.049757973421550236,
80
+ "scr_metric_threshold_5": 0.14686571756372668,
81
+ "scr_dir2_threshold_5": 0.14686571756372668,
82
+ "scr_dir1_threshold_10": -0.15464026775540324,
83
+ "scr_metric_threshold_10": 0.23765912996274793,
84
+ "scr_dir2_threshold_10": 0.23765912996274793,
85
+ "scr_dir1_threshold_20": -1.1812038785157903,
86
+ "scr_metric_threshold_20": 0.21590020651956618,
87
+ "scr_dir2_threshold_20": 0.21590020651956618,
88
+ "scr_dir1_threshold_50": -1.3690589590619218,
89
+ "scr_metric_threshold_50": 0.3159386512624416,
90
+ "scr_dir2_threshold_50": 0.3159386512624416,
91
+ "scr_dir1_threshold_100": -1.5796820637772178,
92
+ "scr_metric_threshold_100": 0.22361457410763316,
93
+ "scr_dir2_threshold_100": 0.22361457410763316,
94
+ "scr_dir1_threshold_500": -1.9224092360141434,
95
+ "scr_metric_threshold_500": 0.15064760938983937,
96
+ "scr_dir2_threshold_500": 0.15064760938983937
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": 0.33333298476841,
103
+ "scr_metric_threshold_2": 0.021844523599026822,
104
+ "scr_dir2_threshold_2": 0.021844523599026822,
105
+ "scr_dir1_threshold_5": -0.5614031968980511,
106
+ "scr_metric_threshold_5": 0.0800970687679908,
107
+ "scr_dir2_threshold_5": 0.0800970687679908,
108
+ "scr_dir1_threshold_10": 0.4035085150925126,
109
+ "scr_metric_threshold_10": 0.12135910742333895,
110
+ "scr_dir2_threshold_10": 0.12135910742333895,
111
+ "scr_dir1_threshold_20": -1.8421032268049216,
112
+ "scr_metric_threshold_20": 0.08495137519633726,
113
+ "scr_dir2_threshold_20": 0.08495137519633726,
114
+ "scr_dir1_threshold_50": -2.210524499582768,
115
+ "scr_metric_threshold_50": 0.29854360680146863,
116
+ "scr_dir2_threshold_50": 0.29854360680146863,
117
+ "scr_dir1_threshold_100": -3.105260681249229,
118
+ "scr_metric_threshold_100": 0.34708739444227193,
119
+ "scr_dir2_threshold_100": 0.34708739444227193,
120
+ "scr_dir1_threshold_500": -3.350874514536203,
121
+ "scr_metric_threshold_500": -0.05097094085497655,
122
+ "scr_dir2_threshold_500": -0.05097094085497655
123
+ },
124
+ {
125
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results",
126
+ "scr_dir1_threshold_2": 0.07767007146192854,
127
+ "scr_metric_threshold_2": 0.1346704699783852,
128
+ "scr_dir2_threshold_2": 0.1346704699783852,
129
+ "scr_dir1_threshold_5": -0.7475731357776643,
130
+ "scr_metric_threshold_5": 0.19484230215355525,
131
+ "scr_dir2_threshold_5": 0.19484230215355525,
132
+ "scr_dir1_threshold_10": -2.097087878670388,
133
+ "scr_metric_threshold_10": 0.42693409399567706,
134
+ "scr_dir2_threshold_10": 0.42693409399567706,
135
+ "scr_dir1_threshold_20": -2.4660198500856167,
136
+ "scr_metric_threshold_20": 0.08882510185734234,
137
+ "scr_dir2_threshold_20": 0.08882510185734234,
138
+ "scr_dir1_threshold_50": -2.893204664440269,
139
+ "scr_metric_threshold_50": -0.12320929873508346,
140
+ "scr_dir2_threshold_50": -0.12320929873508346,
141
+ "scr_dir1_threshold_100": -2.9029132787015213,
142
+ "scr_metric_threshold_100": -0.1719197888799928,
143
+ "scr_dir2_threshold_100": -0.1719197888799928,
144
+ "scr_dir1_threshold_500": -3.194175178654821,
145
+ "scr_metric_threshold_500": -0.13180517716755977,
146
+ "scr_dir2_threshold_500": -0.13180517716755977
147
+ },
148
+ {
149
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results",
150
+ "scr_dir1_threshold_2": 0.6250005321843828,
151
+ "scr_metric_threshold_2": 0.05955332308468681,
152
+ "scr_dir2_threshold_2": 0.05955332308468681,
153
+ "scr_dir1_threshold_5": 0.6250005321843828,
154
+ "scr_metric_threshold_5": -0.02977673549351611,
155
+ "scr_dir2_threshold_5": -0.02977673549351611,
156
+ "scr_dir1_threshold_10": 0.2678569147781217,
157
+ "scr_metric_threshold_10": 0.1538460969606364,
158
+ "scr_dir2_threshold_10": 0.1538460969606364,
159
+ "scr_dir1_threshold_20": -4.785714893925009,
160
+ "scr_metric_threshold_20": 0.3052108362723994,
161
+ "scr_dir2_threshold_20": 0.3052108362723994,
162
+ "scr_dir1_threshold_50": -5.125000532184383,
163
+ "scr_metric_threshold_50": 0.5285359827179067,
164
+ "scr_dir2_threshold_50": 0.5285359827179067,
165
+ "scr_dir1_threshold_100": -5.946429255665635,
166
+ "scr_metric_threshold_100": 0.12406936146712028,
167
+ "scr_dir2_threshold_100": 0.12406936146712028,
168
+ "scr_dir1_threshold_500": -6.357143617406261,
169
+ "scr_metric_threshold_500": -0.15136488721410846,
170
+ "scr_dir2_threshold_500": -0.15136488721410846
171
+ },
172
+ {
173
+ "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results",
174
+ "scr_dir1_threshold_2": 0.3278690847265514,
175
+ "scr_metric_threshold_2": -0.11014488246656093,
176
+ "scr_dir2_threshold_2": -0.11014488246656093,
177
+ "scr_dir1_threshold_5": -0.7868855102060758,
178
+ "scr_metric_threshold_5": -0.19710148683342144,
179
+ "scr_dir2_threshold_5": -0.19710148683342144,
180
+ "scr_dir1_threshold_10": -1.3442628076723893,
181
+ "scr_metric_threshold_10": -0.26666666666666666,
182
+ "scr_dir2_threshold_10": -0.26666666666666666,
183
+ "scr_dir1_threshold_20": -2.229508701302508,
184
+ "scr_metric_threshold_20": -0.21449273859996476,
185
+ "scr_dir2_threshold_20": -0.21449273859996476,
186
+ "scr_dir1_threshold_50": -2.4590169140422704,
187
+ "scr_metric_threshold_50": -0.12173910789994713,
188
+ "scr_dir2_threshold_50": -0.12173910789994713,
189
+ "scr_dir1_threshold_100": -2.385246382192865,
190
+ "scr_metric_threshold_100": -0.26086964033350957,
191
+ "scr_dir2_threshold_100": -0.26086964033350957,
192
+ "scr_dir1_threshold_500": -2.631148317878465,
193
+ "scr_metric_threshold_500": -0.34492755876671954,
194
+ "scr_dir2_threshold_500": -0.34492755876671954
195
+ },
196
+ {
197
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results",
198
+ "scr_dir1_threshold_2": 0.3111110669595241,
199
+ "scr_metric_threshold_2": 0.3122531590459111,
200
+ "scr_dir2_threshold_2": 0.3122531590459111,
201
+ "scr_dir1_threshold_5": 0.3888888336994052,
202
+ "scr_metric_threshold_5": 0.4505929365910292,
203
+ "scr_dir2_threshold_5": 0.4505929365910292,
204
+ "scr_dir1_threshold_10": 0.3888888336994052,
205
+ "scr_metric_threshold_10": 0.5731226423184461,
206
+ "scr_dir2_threshold_10": 0.5731226423184461,
207
+ "scr_dir1_threshold_20": 0.5166666169961314,
208
+ "scr_metric_threshold_20": 0.6363636363636364,
209
+ "scr_dir2_threshold_20": 0.6363636363636364,
210
+ "scr_dir1_threshold_50": 0.2833333167764882,
211
+ "scr_metric_threshold_50": 0.7272727272727273,
212
+ "scr_dir2_threshold_50": 0.7272727272727273,
213
+ "scr_dir1_threshold_100": 0.3000000993410707,
214
+ "scr_metric_threshold_100": 0.6521739437727994,
215
+ "scr_dir2_threshold_100": 0.6521739437727994,
216
+ "scr_dir1_threshold_500": -0.2777778329672615,
217
+ "scr_metric_threshold_500": 0.6877470765455507,
218
+ "scr_dir2_threshold_500": 0.6877470765455507
219
+ },
220
+ {
221
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results",
222
+ "scr_dir1_threshold_2": 0.1853932433771863,
223
+ "scr_metric_threshold_2": 0.0671937475910773,
224
+ "scr_dir2_threshold_2": 0.0671937475910773,
225
+ "scr_dir1_threshold_5": 0.31460675662281373,
226
+ "scr_metric_threshold_5": 0.1501975669998558,
227
+ "scr_dir2_threshold_5": 0.1501975669998558,
228
+ "scr_dir1_threshold_10": 0.4213483108442966,
229
+ "scr_metric_threshold_10": 0.19762848922734483,
230
+ "scr_dir2_threshold_10": 0.19762848922734483,
231
+ "scr_dir1_threshold_20": 0.5056179060416348,
232
+ "scr_metric_threshold_20": 0.2806323086361233,
233
+ "scr_dir2_threshold_20": 0.2806323086361233,
234
+ "scr_dir1_threshold_50": 0.5393260120066543,
235
+ "scr_metric_threshold_50": 0.38735170695437726,
236
+ "scr_dir2_threshold_50": 0.38735170695437726,
237
+ "scr_dir1_threshold_100": 0.5,
238
+ "scr_metric_threshold_100": 0.2569169653181097,
239
+ "scr_dir2_threshold_100": 0.2569169653181097,
240
+ "scr_dir1_threshold_500": 0.5280898650657794,
241
+ "scr_metric_threshold_500": 0.569169888772559,
242
+ "scr_dir2_threshold_500": 0.569169888772559
243
+ },
244
+ {
245
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results",
246
+ "scr_dir1_threshold_2": 0.059139667790609515,
247
+ "scr_metric_threshold_2": 0.33870960764285285,
248
+ "scr_dir2_threshold_2": 0.33870960764285285,
249
+ "scr_dir1_threshold_5": 0.2096776157628017,
250
+ "scr_metric_threshold_5": 0.427419335456348,
251
+ "scr_dir2_threshold_5": 0.427419335456348,
252
+ "scr_dir1_threshold_10": 0.4784945788611154,
253
+ "scr_metric_threshold_10": 0.5201612690019828,
254
+ "scr_dir2_threshold_10": 0.5201612690019828,
255
+ "scr_dir1_threshold_20": 0.4946237248290532,
256
+ "scr_metric_threshold_20": 0.5322581265396864,
257
+ "scr_dir2_threshold_20": 0.5322581265396864,
258
+ "scr_dir1_threshold_50": 0.5698925385876006,
259
+ "scr_metric_threshold_50": 0.7096773418253921,
260
+ "scr_dir2_threshold_50": 0.7096773418253921,
261
+ "scr_dir1_threshold_100": 0.48924744965810646,
262
+ "scr_metric_threshold_100": 0.7741935949047648,
263
+ "scr_dir2_threshold_100": 0.7741935949047648,
264
+ "scr_dir1_threshold_500": -0.5,
265
+ "scr_metric_threshold_500": 0.5241934747341225,
266
+ "scr_dir2_threshold_500": 0.5241934747341225
267
+ },
268
+ {
269
+ "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results",
270
+ "scr_dir1_threshold_2": 0.07981201870298614,
271
+ "scr_metric_threshold_2": 0.049327510076657435,
272
+ "scr_dir2_threshold_2": 0.049327510076657435,
273
+ "scr_dir1_threshold_5": 0.15962431723998594,
274
+ "scr_metric_threshold_5": 0.09865475286797173,
275
+ "scr_dir2_threshold_5": 0.09865475286797173,
276
+ "scr_dir1_threshold_10": 0.24413139102410014,
277
+ "scr_metric_threshold_10": 0.17488800744122396,
278
+ "scr_dir2_threshold_10": 0.17488800744122396,
279
+ "scr_dir1_threshold_20": 0.3568073961249145,
280
+ "scr_metric_threshold_20": 0.013453005890968962,
281
+ "scr_dir2_threshold_20": 0.013453005890968962,
282
+ "scr_dir1_threshold_50": 0.3427230703835713,
283
+ "scr_metric_threshold_50": 0.12107625116269125,
284
+ "scr_dir2_threshold_50": 0.12107625116269125,
285
+ "scr_dir1_threshold_100": 0.41314553859232855,
286
+ "scr_metric_threshold_100": 0.06726476216950167,
287
+ "scr_dir2_threshold_100": 0.06726476216950167,
288
+ "scr_dir1_threshold_500": 0.4037557082640861,
289
+ "scr_metric_threshold_500": 0.10313899906984701,
290
+ "scr_dir2_threshold_500": 0.10313899906984701
291
+ }
292
+ ],
293
+ "sae_bench_commit_hash": "Unknown",
294
+ "sae_lens_id": "blocks.10.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.10.hook_resid_post",
307
+ "hook_layer": 10,
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
+ }