Upload scr/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.hook_resid_post_eval_results.json with huggingface_hub
Browse files
scr/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.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": 1745604160384,
|
74 |
+
"eval_result_metrics": {
|
75 |
+
"scr_metrics": {
|
76 |
+
"scr_dir1_threshold_2": 0.2744988192668235,
|
77 |
+
"scr_metric_threshold_2": 0.10396997212912865,
|
78 |
+
"scr_dir2_threshold_2": 0.10396997212912865,
|
79 |
+
"scr_dir1_threshold_5": -0.701007737215005,
|
80 |
+
"scr_metric_threshold_5": 0.25939921292472773,
|
81 |
+
"scr_dir2_threshold_5": 0.25939921292472773,
|
82 |
+
"scr_dir1_threshold_10": -0.4687670318047186,
|
83 |
+
"scr_metric_threshold_10": 0.3781358757326926,
|
84 |
+
"scr_dir2_threshold_10": 0.3781358757326926,
|
85 |
+
"scr_dir1_threshold_20": -0.5259256026188907,
|
86 |
+
"scr_metric_threshold_20": 0.4359074813736896,
|
87 |
+
"scr_dir2_threshold_20": 0.4359074813736896,
|
88 |
+
"scr_dir1_threshold_50": -0.8744274975284453,
|
89 |
+
"scr_metric_threshold_50": 0.40439074477982234,
|
90 |
+
"scr_dir2_threshold_50": 0.40439074477982234,
|
91 |
+
"scr_dir1_threshold_100": -1.3161551352158563,
|
92 |
+
"scr_metric_threshold_100": 0.37081197663713317,
|
93 |
+
"scr_dir2_threshold_100": 0.37081197663713317,
|
94 |
+
"scr_dir1_threshold_500": -2.11771227264383,
|
95 |
+
"scr_metric_threshold_500": 0.24836667106819627,
|
96 |
+
"scr_dir2_threshold_500": 0.24836667106819627
|
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": 0.33928527713472234,
|
103 |
+
"scr_metric_threshold_2": 0.01694907915872299,
|
104 |
+
"scr_dir2_threshold_2": 0.01694907915872299,
|
105 |
+
"scr_dir1_threshold_5": -2.3571400441712567,
|
106 |
+
"scr_metric_threshold_5": 0.08716710516251286,
|
107 |
+
"scr_dir2_threshold_5": 0.08716710516251286,
|
108 |
+
"scr_dir1_threshold_10": -1.4107128602219206,
|
109 |
+
"scr_metric_threshold_10": 0.15980626325048816,
|
110 |
+
"scr_dir2_threshold_10": 0.15980626325048816,
|
111 |
+
"scr_dir1_threshold_20": -1.4999984034485512,
|
112 |
+
"scr_metric_threshold_20": 0.2203390393116457,
|
113 |
+
"scr_dir2_threshold_20": 0.2203390393116457,
|
114 |
+
"scr_dir1_threshold_50": -3.214282749261595,
|
115 |
+
"scr_metric_threshold_50": 0.21791761858510403,
|
116 |
+
"scr_dir2_threshold_50": 0.21791761858510403,
|
117 |
+
"scr_dir1_threshold_100": -3.392853835714856,
|
118 |
+
"scr_metric_threshold_100": 0.3728813290248651,
|
119 |
+
"scr_dir2_threshold_100": 0.3728813290248651,
|
120 |
+
"scr_dir1_threshold_500": -5.928566030707006,
|
121 |
+
"scr_metric_threshold_500": 0.14043590768640163,
|
122 |
+
"scr_dir2_threshold_500": 0.14043590768640163
|
123 |
+
},
|
124 |
+
{
|
125 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results",
|
126 |
+
"scr_dir1_threshold_2": 0.4392522791814889,
|
127 |
+
"scr_metric_threshold_2": 0.2052024171921838,
|
128 |
+
"scr_dir2_threshold_2": 0.2052024171921838,
|
129 |
+
"scr_dir1_threshold_5": -1.2710284642834462,
|
130 |
+
"scr_metric_threshold_5": 0.46242780043194415,
|
131 |
+
"scr_dir2_threshold_5": 0.46242780043194415,
|
132 |
+
"scr_dir1_threshold_10": -1.1682243719508323,
|
133 |
+
"scr_metric_threshold_10": 0.6242774875161867,
|
134 |
+
"scr_dir2_threshold_10": 0.6242774875161867,
|
135 |
+
"scr_dir1_threshold_20": -1.2429908832740444,
|
136 |
+
"scr_metric_threshold_20": 0.5635839409934715,
|
137 |
+
"scr_dir2_threshold_20": 0.5635839409934715,
|
138 |
+
"scr_dir1_threshold_50": -1.0093462317049937,
|
139 |
+
"scr_metric_threshold_50": 0.19364170064791625,
|
140 |
+
"scr_dir2_threshold_50": 0.19364170064791625,
|
141 |
+
"scr_dir1_threshold_100": -1.8878507900679715,
|
142 |
+
"scr_metric_threshold_100": 0.11849712924405292,
|
143 |
+
"scr_dir2_threshold_100": 0.11849712924405292,
|
144 |
+
"scr_dir1_threshold_500": -2.4672904172436807,
|
145 |
+
"scr_metric_threshold_500": -0.3150287936933465,
|
146 |
+
"scr_dir2_threshold_500": -0.3150287936933465
|
147 |
+
},
|
148 |
+
{
|
149 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results",
|
150 |
+
"scr_dir1_threshold_2": 0.6440684301554974,
|
151 |
+
"scr_metric_threshold_2": 0.06015049050583274,
|
152 |
+
"scr_dir2_threshold_2": 0.06015049050583274,
|
153 |
+
"scr_dir1_threshold_5": -2.338982639899137,
|
154 |
+
"scr_metric_threshold_5": 0.14786969440299008,
|
155 |
+
"scr_dir2_threshold_5": 0.14786969440299008,
|
156 |
+
"scr_dir1_threshold_10": -2.152541379756247,
|
157 |
+
"scr_metric_threshold_10": 0.2982456218974591,
|
158 |
+
"scr_dir2_threshold_10": 0.2982456218974591,
|
159 |
+
"scr_dir1_threshold_20": -2.5593207496848027,
|
160 |
+
"scr_metric_threshold_20": 0.3734336603372218,
|
161 |
+
"scr_dir2_threshold_20": 0.3734336603372218,
|
162 |
+
"scr_dir1_threshold_50": -3.661016349852908,
|
163 |
+
"scr_metric_threshold_50": 0.09523820194170696,
|
164 |
+
"scr_dir2_threshold_50": 0.09523820194170696,
|
165 |
+
"scr_dir1_threshold_100": -5.949151189667992,
|
166 |
+
"scr_metric_threshold_100": 0.07518803843976271,
|
167 |
+
"scr_dir2_threshold_100": 0.07518803843976271,
|
168 |
+
"scr_dir1_threshold_500": -6.322032699705816,
|
169 |
+
"scr_metric_threshold_500": -0.1578946267689058,
|
170 |
+
"scr_dir2_threshold_500": -0.1578946267689058
|
171 |
+
},
|
172 |
+
{
|
173 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results",
|
174 |
+
"scr_dir1_threshold_2": 0.0852711387397361,
|
175 |
+
"scr_metric_threshold_2": 0.023809570270056916,
|
176 |
+
"scr_dir2_threshold_2": 0.023809570270056916,
|
177 |
+
"scr_dir1_threshold_5": -0.6046511950268475,
|
178 |
+
"scr_metric_threshold_5": 0.18452390350079695,
|
179 |
+
"scr_dir2_threshold_5": 0.18452390350079695,
|
180 |
+
"scr_dir1_threshold_10": -0.13953492670246334,
|
181 |
+
"scr_metric_threshold_10": 0.3511905406016698,
|
182 |
+
"scr_dir2_threshold_10": 0.3511905406016698,
|
183 |
+
"scr_dir1_threshold_20": -0.4728681060186364,
|
184 |
+
"scr_metric_threshold_20": 0.4404761630222391,
|
185 |
+
"scr_dir2_threshold_20": 0.4404761630222391,
|
186 |
+
"scr_dir1_threshold_50": -0.8527132356032845,
|
187 |
+
"scr_metric_threshold_50": 0.5565476850426945,
|
188 |
+
"scr_dir2_threshold_50": 0.5565476850426945,
|
189 |
+
"scr_dir1_threshold_100": -0.3255813416219208,
|
190 |
+
"scr_metric_threshold_100": 0.16666663710087287,
|
191 |
+
"scr_dir2_threshold_100": 0.16666663710087287,
|
192 |
+
"scr_dir1_threshold_500": -1.8837210484317741,
|
193 |
+
"scr_metric_threshold_500": -0.27678567785643265,
|
194 |
+
"scr_dir2_threshold_500": -0.27678567785643265
|
195 |
+
},
|
196 |
+
{
|
197 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results",
|
198 |
+
"scr_dir1_threshold_2": 0.3199999046325532,
|
199 |
+
"scr_metric_threshold_2": 0.2956204951187285,
|
200 |
+
"scr_dir2_threshold_2": 0.2956204951187285,
|
201 |
+
"scr_dir1_threshold_5": 0.31333332273695036,
|
202 |
+
"scr_metric_threshold_5": 0.5620438559586579,
|
203 |
+
"scr_dir2_threshold_5": 0.5620438559586579,
|
204 |
+
"scr_dir1_threshold_10": 0.3199999046325532,
|
205 |
+
"scr_metric_threshold_10": 0.7043795048812714,
|
206 |
+
"scr_dir2_threshold_10": 0.7043795048812714,
|
207 |
+
"scr_dir1_threshold_20": 0.5733331955803547,
|
208 |
+
"scr_metric_threshold_20": 0.7700729482409889,
|
209 |
+
"scr_dir2_threshold_20": 0.7700729482409889,
|
210 |
+
"scr_dir1_threshold_50": 0.5600000317891489,
|
211 |
+
"scr_metric_threshold_50": 0.7080293098174923,
|
212 |
+
"scr_dir2_threshold_50": 0.7080293098174923,
|
213 |
+
"scr_dir1_threshold_100": 0.1399998092651064,
|
214 |
+
"scr_metric_threshold_100": 0.7554743811015894,
|
215 |
+
"scr_dir2_threshold_100": 0.7554743811015894,
|
216 |
+
"scr_dir1_threshold_500": -0.3933334975772695,
|
217 |
+
"scr_metric_threshold_500": 0.7737225356420485,
|
218 |
+
"scr_dir2_threshold_500": 0.7737225356420485
|
219 |
+
},
|
220 |
+
{
|
221 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results",
|
222 |
+
"scr_dir1_threshold_2": 0.24793366445932427,
|
223 |
+
"scr_metric_threshold_2": 0.046822733807671976,
|
224 |
+
"scr_dir2_threshold_2": 0.046822733807671976,
|
225 |
+
"scr_dir1_threshold_5": 0.38842970321318315,
|
226 |
+
"scr_metric_threshold_5": 0.09699010523109457,
|
227 |
+
"scr_dir2_threshold_5": 0.09699010523109457,
|
228 |
+
"scr_dir1_threshold_10": 0.46280990107106107,
|
229 |
+
"scr_metric_threshold_10": 0.15719079146190446,
|
230 |
+
"scr_dir2_threshold_10": 0.15719079146190446,
|
231 |
+
"scr_dir1_threshold_20": 0.47933910759525783,
|
232 |
+
"scr_metric_threshold_20": 0.2608695738846357,
|
233 |
+
"scr_dir2_threshold_20": 0.2608695738846357,
|
234 |
+
"scr_dir1_threshold_50": 0.5619836624150326,
|
235 |
+
"scr_metric_threshold_50": 0.4247492418847984,
|
236 |
+
"scr_dir2_threshold_50": 0.4247492418847984,
|
237 |
+
"scr_dir1_threshold_100": 0.4380163375849674,
|
238 |
+
"scr_metric_threshold_100": 0.5317726619232802,
|
239 |
+
"scr_dir2_threshold_100": 0.5317726619232802,
|
240 |
+
"scr_dir1_threshold_500": 0.3057851483934084,
|
241 |
+
"scr_metric_threshold_500": 0.665551986808038,
|
242 |
+
"scr_dir2_threshold_500": 0.665551986808038
|
243 |
+
},
|
244 |
+
{
|
245 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results",
|
246 |
+
"scr_dir1_threshold_2": 0.06400012969970703,
|
247 |
+
"scr_metric_threshold_2": 0.12582795181103357,
|
248 |
+
"scr_dir2_threshold_2": 0.12582795181103357,
|
249 |
+
"scr_dir1_threshold_5": 0.14400005340576172,
|
250 |
+
"scr_metric_threshold_5": 0.3907285186471721,
|
251 |
+
"scr_dir2_threshold_5": 0.3907285186471721,
|
252 |
+
"scr_dir1_threshold_10": 0.1919999122619629,
|
253 |
+
"scr_metric_threshold_10": 0.5496688192755865,
|
254 |
+
"scr_dir2_threshold_10": 0.5496688192755865,
|
255 |
+
"scr_dir1_threshold_20": 0.29600000381469727,
|
256 |
+
"scr_metric_threshold_20": 0.6125828938642751,
|
257 |
+
"scr_dir2_threshold_20": 0.6125828938642751,
|
258 |
+
"scr_dir1_threshold_50": 0.3839998245239258,
|
259 |
+
"scr_metric_threshold_50": 0.715231747560552,
|
260 |
+
"scr_dir2_threshold_50": 0.715231747560552,
|
261 |
+
"scr_dir1_threshold_100": 0.3920001983642578,
|
262 |
+
"scr_metric_threshold_100": 0.7615893516910348,
|
263 |
+
"scr_dir2_threshold_100": 0.7615893516910348,
|
264 |
+
"scr_dir1_threshold_500": 0.4159998893737793,
|
265 |
+
"scr_metric_threshold_500": 0.7880794675845517,
|
266 |
+
"scr_dir2_threshold_500": 0.7880794675845517
|
267 |
+
},
|
268 |
+
{
|
269 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results",
|
270 |
+
"scr_dir1_threshold_2": 0.05617973013155886,
|
271 |
+
"scr_metric_threshold_2": 0.05737703916879863,
|
272 |
+
"scr_dir2_threshold_2": 0.05737703916879863,
|
273 |
+
"scr_dir1_threshold_5": 0.11797736630475257,
|
274 |
+
"scr_metric_threshold_5": 0.14344272006265327,
|
275 |
+
"scr_dir2_threshold_5": 0.14344272006265327,
|
276 |
+
"scr_dir1_threshold_10": 0.14606756622813716,
|
277 |
+
"scr_metric_threshold_10": 0.18032797697697478,
|
278 |
+
"scr_dir2_threshold_10": 0.18032797697697478,
|
279 |
+
"scr_dir1_threshold_20": 0.21910101448460056,
|
280 |
+
"scr_metric_threshold_20": 0.2459016313350389,
|
281 |
+
"scr_dir2_threshold_20": 0.2459016313350389,
|
282 |
+
"scr_dir1_threshold_50": 0.23595506746711029,
|
283 |
+
"scr_metric_threshold_50": 0.3237704527583147,
|
284 |
+
"scr_dir2_threshold_50": 0.3237704527583147,
|
285 |
+
"scr_dir1_threshold_100": 0.05617973013155886,
|
286 |
+
"scr_metric_threshold_100": 0.18442628457160753,
|
287 |
+
"scr_dir2_threshold_100": 0.18442628457160753,
|
288 |
+
"scr_dir1_threshold_500": -0.6685395252522818,
|
289 |
+
"scr_metric_threshold_500": 0.36885256914321507,
|
290 |
+
"scr_dir2_threshold_500": 0.36885256914321507
|
291 |
+
}
|
292 |
+
],
|
293 |
+
"sae_bench_commit_hash": "Unknown",
|
294 |
+
"sae_lens_id": "blocks.8.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.8.hook_resid_post",
|
307 |
+
"hook_layer": 8,
|
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 |
+
}
|