chanind commited on
Commit
337ab66
·
verified ·
1 Parent(s): 4d8805c

Upload absorption/gemma-2-2b-res-matryoshka-dc_blocks.20.hook_resid_post_eval_results.json with huggingface_hub

Browse files
absorption/gemma-2-2b-res-matryoshka-dc_blocks.20.hook_resid_post_eval_results.json ADDED
@@ -0,0 +1,275 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "eval_type_id": "absorption_first_letter",
3
+ "eval_config": {
4
+ "model_name": "gemma-2-2b",
5
+ "random_seed": 42,
6
+ "f1_jump_threshold": 0.03,
7
+ "max_k_value": 10,
8
+ "prompt_template": "{word} has the first letter:",
9
+ "prompt_token_pos": -6,
10
+ "llm_batch_size": 32,
11
+ "llm_dtype": "bfloat16",
12
+ "k_sparse_probe_l1_decay": 0.01,
13
+ "k_sparse_probe_batch_size": 4096,
14
+ "k_sparse_probe_num_epochs": 50,
15
+ "min_GT_probe_f1": 0.6,
16
+ "min_feats_for_eval": 20
17
+ },
18
+ "eval_id": "a15aaf76-1ebc-4334-8744-86b1c8367c42",
19
+ "datetime_epoch_millis": 1745797579234,
20
+ "eval_result_metrics": {
21
+ "mean": {
22
+ "mean_absorption_fraction_score": 0.21342511542943934,
23
+ "mean_full_absorption_score": 0.18598426121803746,
24
+ "mean_num_split_features": 1.2692307692307692,
25
+ "std_dev_absorption_fraction_score": 0.2553427420423028,
26
+ "std_dev_full_absorption_score": 0.2420585309725433,
27
+ "std_dev_num_split_features": 0.533493565673837
28
+ }
29
+ },
30
+ "eval_result_details": [
31
+ {
32
+ "first_letter": "a",
33
+ "mean_absorption_fraction": 0.33082372413170374,
34
+ "full_absorption_rate": 0.25265225933202357,
35
+ "num_full_absorption": 643,
36
+ "num_probe_true_positives": 2545,
37
+ "num_split_features": 1
38
+ },
39
+ {
40
+ "first_letter": "b",
41
+ "mean_absorption_fraction": 0.1524404219476543,
42
+ "full_absorption_rate": 0.11090225563909774,
43
+ "num_full_absorption": 177,
44
+ "num_probe_true_positives": 1596,
45
+ "num_split_features": 1
46
+ },
47
+ {
48
+ "first_letter": "c",
49
+ "mean_absorption_fraction": 0.5049150717161012,
50
+ "full_absorption_rate": 0.46023564064801176,
51
+ "num_full_absorption": 1250,
52
+ "num_probe_true_positives": 2716,
53
+ "num_split_features": 2
54
+ },
55
+ {
56
+ "first_letter": "d",
57
+ "mean_absorption_fraction": 0.17806306279643594,
58
+ "full_absorption_rate": 0.15127092374457532,
59
+ "num_full_absorption": 244,
60
+ "num_probe_true_positives": 1613,
61
+ "num_split_features": 2
62
+ },
63
+ {
64
+ "first_letter": "e",
65
+ "mean_absorption_fraction": 0.15561198492359832,
66
+ "full_absorption_rate": 0.15040397762585456,
67
+ "num_full_absorption": 242,
68
+ "num_probe_true_positives": 1609,
69
+ "num_split_features": 1
70
+ },
71
+ {
72
+ "first_letter": "f",
73
+ "mean_absorption_fraction": 0.1774656475503202,
74
+ "full_absorption_rate": 0.15103734439834024,
75
+ "num_full_absorption": 182,
76
+ "num_probe_true_positives": 1205,
77
+ "num_split_features": 1
78
+ },
79
+ {
80
+ "first_letter": "g",
81
+ "mean_absorption_fraction": 0.04563388313001719,
82
+ "full_absorption_rate": 0.028745644599303136,
83
+ "num_full_absorption": 33,
84
+ "num_probe_true_positives": 1148,
85
+ "num_split_features": 2
86
+ },
87
+ {
88
+ "first_letter": "h",
89
+ "mean_absorption_fraction": 0.04946583375206639,
90
+ "full_absorption_rate": 0.048503611971104234,
91
+ "num_full_absorption": 47,
92
+ "num_probe_true_positives": 969,
93
+ "num_split_features": 1
94
+ },
95
+ {
96
+ "first_letter": "i",
97
+ "mean_absorption_fraction": 0.17653161862457492,
98
+ "full_absorption_rate": 0.18044659022329512,
99
+ "num_full_absorption": 299,
100
+ "num_probe_true_positives": 1657,
101
+ "num_split_features": 1
102
+ },
103
+ {
104
+ "first_letter": "j",
105
+ "mean_absorption_fraction": 3.072431234371448e-05,
106
+ "full_absorption_rate": 0.0,
107
+ "num_full_absorption": 0,
108
+ "num_probe_true_positives": 415,
109
+ "num_split_features": 1
110
+ },
111
+ {
112
+ "first_letter": "k",
113
+ "mean_absorption_fraction": 0.0036572058006725935,
114
+ "full_absorption_rate": 0.0030627871362940277,
115
+ "num_full_absorption": 2,
116
+ "num_probe_true_positives": 653,
117
+ "num_split_features": 1
118
+ },
119
+ {
120
+ "first_letter": "l",
121
+ "mean_absorption_fraction": 0.1496291794186504,
122
+ "full_absorption_rate": 0.14156378600823044,
123
+ "num_full_absorption": 172,
124
+ "num_probe_true_positives": 1215,
125
+ "num_split_features": 1
126
+ },
127
+ {
128
+ "first_letter": "m",
129
+ "mean_absorption_fraction": 0.3325867333915395,
130
+ "full_absorption_rate": 0.2749445676274945,
131
+ "num_full_absorption": 496,
132
+ "num_probe_true_positives": 1804,
133
+ "num_split_features": 1
134
+ },
135
+ {
136
+ "first_letter": "n",
137
+ "mean_absorption_fraction": 0.04191789764464963,
138
+ "full_absorption_rate": 0.04187192118226601,
139
+ "num_full_absorption": 34,
140
+ "num_probe_true_positives": 812,
141
+ "num_split_features": 1
142
+ },
143
+ {
144
+ "first_letter": "o",
145
+ "mean_absorption_fraction": 0.06027940422095779,
146
+ "full_absorption_rate": 0.08527131782945736,
147
+ "num_full_absorption": 88,
148
+ "num_probe_true_positives": 1032,
149
+ "num_split_features": 1
150
+ },
151
+ {
152
+ "first_letter": "p",
153
+ "mean_absorption_fraction": 0.17716080390336,
154
+ "full_absorption_rate": 0.16330904414916417,
155
+ "num_full_absorption": 381,
156
+ "num_probe_true_positives": 2333,
157
+ "num_split_features": 2
158
+ },
159
+ {
160
+ "first_letter": "q",
161
+ "mean_absorption_fraction": 0.0,
162
+ "full_absorption_rate": 0.0,
163
+ "num_full_absorption": 0,
164
+ "num_probe_true_positives": 199,
165
+ "num_split_features": 1
166
+ },
167
+ {
168
+ "first_letter": "r",
169
+ "mean_absorption_fraction": 0.027809109885836673,
170
+ "full_absorption_rate": 0.03371501272264631,
171
+ "num_full_absorption": 53,
172
+ "num_probe_true_positives": 1572,
173
+ "num_split_features": 3
174
+ },
175
+ {
176
+ "first_letter": "s",
177
+ "mean_absorption_fraction": 0.8186751008065684,
178
+ "full_absorption_rate": 0.8231292517006803,
179
+ "num_full_absorption": 2299,
180
+ "num_probe_true_positives": 2793,
181
+ "num_split_features": 1
182
+ },
183
+ {
184
+ "first_letter": "t",
185
+ "mean_absorption_fraction": 0.2060006947973954,
186
+ "full_absorption_rate": 0.18070067609096496,
187
+ "num_full_absorption": 294,
188
+ "num_probe_true_positives": 1627,
189
+ "num_split_features": 1
190
+ },
191
+ {
192
+ "first_letter": "u",
193
+ "mean_absorption_fraction": 0.20808960266963086,
194
+ "full_absorption_rate": 0.09986859395532194,
195
+ "num_full_absorption": 76,
196
+ "num_probe_true_positives": 761,
197
+ "num_split_features": 1
198
+ },
199
+ {
200
+ "first_letter": "v",
201
+ "mean_absorption_fraction": 0.0,
202
+ "full_absorption_rate": 0.0024125452352231603,
203
+ "num_full_absorption": 2,
204
+ "num_probe_true_positives": 829,
205
+ "num_split_features": 2
206
+ },
207
+ {
208
+ "first_letter": "w",
209
+ "mean_absorption_fraction": 0.08197207883170882,
210
+ "full_absorption_rate": 0.054380664652567974,
211
+ "num_full_absorption": 36,
212
+ "num_probe_true_positives": 662,
213
+ "num_split_features": 1
214
+ },
215
+ {
216
+ "first_letter": "x",
217
+ "mean_absorption_fraction": 0.9454545454545454,
218
+ "full_absorption_rate": 0.9636363636363636,
219
+ "num_full_absorption": 106,
220
+ "num_probe_true_positives": 110,
221
+ "num_split_features": 1
222
+ },
223
+ {
224
+ "first_letter": "y",
225
+ "mean_absorption_fraction": 0.6893315709316808,
226
+ "full_absorption_rate": 0.43352601156069365,
227
+ "num_full_absorption": 75,
228
+ "num_probe_true_positives": 173,
229
+ "num_split_features": 1
230
+ },
231
+ {
232
+ "first_letter": "z",
233
+ "mean_absorption_fraction": 0.03550710052341046,
234
+ "full_absorption_rate": 0.0,
235
+ "num_full_absorption": 0,
236
+ "num_probe_true_positives": 219,
237
+ "num_split_features": 1
238
+ }
239
+ ],
240
+ "sae_bench_commit_hash": "Unknown",
241
+ "sae_lens_id": "blocks.20.hook_resid_post",
242
+ "sae_lens_release_id": "gemma-2-2b-res-matryoshka-dc",
243
+ "sae_lens_version": "5.9.1",
244
+ "sae_cfg_dict": {
245
+ "architecture": "jumprelu",
246
+ "d_in": 2304,
247
+ "d_sae": 32768,
248
+ "activation_fn_str": "relu",
249
+ "apply_b_dec_to_input": true,
250
+ "finetuning_scaling_factor": false,
251
+ "context_size": 1024,
252
+ "model_name": "gemma-2-2b",
253
+ "hook_name": "blocks.20.hook_resid_post",
254
+ "hook_layer": 20,
255
+ "hook_head_index": null,
256
+ "prepend_bos": true,
257
+ "dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B",
258
+ "dataset_trust_remote_code": true,
259
+ "normalize_activations": "none",
260
+ "dtype": "torch.bfloat16",
261
+ "device": "cuda",
262
+ "sae_lens_training_version": "5.5.1",
263
+ "activation_fn_kwargs": {
264
+ "k": 40
265
+ },
266
+ "neuronpedia_id": null,
267
+ "model_from_pretrained_kwargs": {
268
+ "center_writing_weights": false
269
+ },
270
+ "seqpos_slice": [
271
+ null
272
+ ]
273
+ },
274
+ "eval_result_unstructured": null
275
+ }