Upload absorption/gemma-2-2b-res-matryoshka-dc_blocks.18.hook_resid_post_eval_results.json with huggingface_hub
Browse files
absorption/gemma-2-2b-res-matryoshka-dc_blocks.18.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": 1745796671224,
|
| 20 |
+
"eval_result_metrics": {
|
| 21 |
+
"mean": {
|
| 22 |
+
"mean_absorption_fraction_score": 0.1424107673166,
|
| 23 |
+
"mean_full_absorption_score": 0.12510257067657027,
|
| 24 |
+
"mean_num_split_features": 1.4230769230769231,
|
| 25 |
+
"std_dev_absorption_fraction_score": 0.1886807126127507,
|
| 26 |
+
"std_dev_full_absorption_score": 0.1826445961743382,
|
| 27 |
+
"std_dev_num_split_features": 0.986836437230686
|
| 28 |
+
}
|
| 29 |
+
},
|
| 30 |
+
"eval_result_details": [
|
| 31 |
+
{
|
| 32 |
+
"first_letter": "a",
|
| 33 |
+
"mean_absorption_fraction": 0.5328801123530625,
|
| 34 |
+
"full_absorption_rate": 0.4609069585613761,
|
| 35 |
+
"num_full_absorption": 1179,
|
| 36 |
+
"num_probe_true_positives": 2558,
|
| 37 |
+
"num_split_features": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"first_letter": "b",
|
| 41 |
+
"mean_absorption_fraction": 0.08956793828369938,
|
| 42 |
+
"full_absorption_rate": 0.044859813084112146,
|
| 43 |
+
"num_full_absorption": 72,
|
| 44 |
+
"num_probe_true_positives": 1605,
|
| 45 |
+
"num_split_features": 1
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"first_letter": "c",
|
| 49 |
+
"mean_absorption_fraction": 0.7127686456185302,
|
| 50 |
+
"full_absorption_rate": 0.7023853211009174,
|
| 51 |
+
"num_full_absorption": 1914,
|
| 52 |
+
"num_probe_true_positives": 2725,
|
| 53 |
+
"num_split_features": 1
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"first_letter": "d",
|
| 57 |
+
"mean_absorption_fraction": 0.13557929372505884,
|
| 58 |
+
"full_absorption_rate": 0.08487084870848709,
|
| 59 |
+
"num_full_absorption": 138,
|
| 60 |
+
"num_probe_true_positives": 1626,
|
| 61 |
+
"num_split_features": 1
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"first_letter": "e",
|
| 65 |
+
"mean_absorption_fraction": 0.20619154600483394,
|
| 66 |
+
"full_absorption_rate": 0.19260204081632654,
|
| 67 |
+
"num_full_absorption": 302,
|
| 68 |
+
"num_probe_true_positives": 1568,
|
| 69 |
+
"num_split_features": 1
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"first_letter": "f",
|
| 73 |
+
"mean_absorption_fraction": 0.09095697692142744,
|
| 74 |
+
"full_absorption_rate": 0.06448911222780569,
|
| 75 |
+
"num_full_absorption": 77,
|
| 76 |
+
"num_probe_true_positives": 1194,
|
| 77 |
+
"num_split_features": 1
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"first_letter": "g",
|
| 81 |
+
"mean_absorption_fraction": 0.005152931679235823,
|
| 82 |
+
"full_absorption_rate": 0.009769094138543518,
|
| 83 |
+
"num_full_absorption": 11,
|
| 84 |
+
"num_probe_true_positives": 1126,
|
| 85 |
+
"num_split_features": 2
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"first_letter": "h",
|
| 89 |
+
"mean_absorption_fraction": 0.007743445860113119,
|
| 90 |
+
"full_absorption_rate": 0.006060606060606061,
|
| 91 |
+
"num_full_absorption": 6,
|
| 92 |
+
"num_probe_true_positives": 990,
|
| 93 |
+
"num_split_features": 1
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"first_letter": "i",
|
| 97 |
+
"mean_absorption_fraction": 0.06884379827803416,
|
| 98 |
+
"full_absorption_rate": 0.057356608478802994,
|
| 99 |
+
"num_full_absorption": 92,
|
| 100 |
+
"num_probe_true_positives": 1604,
|
| 101 |
+
"num_split_features": 1
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"first_letter": "j",
|
| 105 |
+
"mean_absorption_fraction": 0.0,
|
| 106 |
+
"full_absorption_rate": 0.0,
|
| 107 |
+
"num_full_absorption": 0,
|
| 108 |
+
"num_probe_true_positives": 455,
|
| 109 |
+
"num_split_features": 1
|
| 110 |
+
},
|
| 111 |
+
{
|
| 112 |
+
"first_letter": "k",
|
| 113 |
+
"mean_absorption_fraction": 0.005084256830246763,
|
| 114 |
+
"full_absorption_rate": 0.008849557522123894,
|
| 115 |
+
"num_full_absorption": 6,
|
| 116 |
+
"num_probe_true_positives": 678,
|
| 117 |
+
"num_split_features": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"first_letter": "l",
|
| 121 |
+
"mean_absorption_fraction": 0.06506943714982681,
|
| 122 |
+
"full_absorption_rate": 0.04991394148020654,
|
| 123 |
+
"num_full_absorption": 58,
|
| 124 |
+
"num_probe_true_positives": 1162,
|
| 125 |
+
"num_split_features": 1
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"first_letter": "m",
|
| 129 |
+
"mean_absorption_fraction": 0.22891553059761788,
|
| 130 |
+
"full_absorption_rate": 0.18057921635434412,
|
| 131 |
+
"num_full_absorption": 318,
|
| 132 |
+
"num_probe_true_positives": 1761,
|
| 133 |
+
"num_split_features": 1
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"first_letter": "n",
|
| 137 |
+
"mean_absorption_fraction": 0.07043036537933946,
|
| 138 |
+
"full_absorption_rate": 0.0364963503649635,
|
| 139 |
+
"num_full_absorption": 30,
|
| 140 |
+
"num_probe_true_positives": 822,
|
| 141 |
+
"num_split_features": 1
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"first_letter": "o",
|
| 145 |
+
"mean_absorption_fraction": 0.09204694481402126,
|
| 146 |
+
"full_absorption_rate": 0.10933081998114987,
|
| 147 |
+
"num_full_absorption": 116,
|
| 148 |
+
"num_probe_true_positives": 1061,
|
| 149 |
+
"num_split_features": 1
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"first_letter": "p",
|
| 153 |
+
"mean_absorption_fraction": 0.397383525427396,
|
| 154 |
+
"full_absorption_rate": 0.3700887198986058,
|
| 155 |
+
"num_full_absorption": 876,
|
| 156 |
+
"num_probe_true_positives": 2367,
|
| 157 |
+
"num_split_features": 1
|
| 158 |
+
},
|
| 159 |
+
{
|
| 160 |
+
"first_letter": "q",
|
| 161 |
+
"mean_absorption_fraction": 0.0009224929378531073,
|
| 162 |
+
"full_absorption_rate": 0.0,
|
| 163 |
+
"num_full_absorption": 0,
|
| 164 |
+
"num_probe_true_positives": 177,
|
| 165 |
+
"num_split_features": 1
|
| 166 |
+
},
|
| 167 |
+
{
|
| 168 |
+
"first_letter": "r",
|
| 169 |
+
"mean_absorption_fraction": 0.12721780451576514,
|
| 170 |
+
"full_absorption_rate": 0.13096695226438188,
|
| 171 |
+
"num_full_absorption": 214,
|
| 172 |
+
"num_probe_true_positives": 1634,
|
| 173 |
+
"num_split_features": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"first_letter": "s",
|
| 177 |
+
"mean_absorption_fraction": 0.5164213398692016,
|
| 178 |
+
"full_absorption_rate": 0.5086206896551724,
|
| 179 |
+
"num_full_absorption": 1416,
|
| 180 |
+
"num_probe_true_positives": 2784,
|
| 181 |
+
"num_split_features": 5
|
| 182 |
+
},
|
| 183 |
+
{
|
| 184 |
+
"first_letter": "t",
|
| 185 |
+
"mean_absorption_fraction": 0.12789020911615354,
|
| 186 |
+
"full_absorption_rate": 0.11675741076830005,
|
| 187 |
+
"num_full_absorption": 193,
|
| 188 |
+
"num_probe_true_positives": 1653,
|
| 189 |
+
"num_split_features": 1
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"first_letter": "u",
|
| 193 |
+
"mean_absorption_fraction": 0.10128583813911718,
|
| 194 |
+
"full_absorption_rate": 0.07472826086956522,
|
| 195 |
+
"num_full_absorption": 55,
|
| 196 |
+
"num_probe_true_positives": 736,
|
| 197 |
+
"num_split_features": 2
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"first_letter": "v",
|
| 201 |
+
"mean_absorption_fraction": 0.02339272024828267,
|
| 202 |
+
"full_absorption_rate": 0.022857142857142857,
|
| 203 |
+
"num_full_absorption": 20,
|
| 204 |
+
"num_probe_true_positives": 875,
|
| 205 |
+
"num_split_features": 1
|
| 206 |
+
},
|
| 207 |
+
{
|
| 208 |
+
"first_letter": "w",
|
| 209 |
+
"mean_absorption_fraction": 0.01206131369374057,
|
| 210 |
+
"full_absorption_rate": 0.009188361408882083,
|
| 211 |
+
"num_full_absorption": 6,
|
| 212 |
+
"num_probe_true_positives": 653,
|
| 213 |
+
"num_split_features": 1
|
| 214 |
+
},
|
| 215 |
+
{
|
| 216 |
+
"first_letter": "x",
|
| 217 |
+
"mean_absorption_fraction": 0.07067388653125427,
|
| 218 |
+
"full_absorption_rate": 0.0,
|
| 219 |
+
"num_full_absorption": 0,
|
| 220 |
+
"num_probe_true_positives": 96,
|
| 221 |
+
"num_split_features": 2
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"first_letter": "y",
|
| 225 |
+
"mean_absorption_fraction": 0.005526205377930516,
|
| 226 |
+
"full_absorption_rate": 0.01098901098901099,
|
| 227 |
+
"num_full_absorption": 2,
|
| 228 |
+
"num_probe_true_positives": 182,
|
| 229 |
+
"num_split_features": 1
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"first_letter": "z",
|
| 233 |
+
"mean_absorption_fraction": 0.008673390879857475,
|
| 234 |
+
"full_absorption_rate": 0.0,
|
| 235 |
+
"num_full_absorption": 0,
|
| 236 |
+
"num_probe_true_positives": 230,
|
| 237 |
+
"num_split_features": 1
|
| 238 |
+
}
|
| 239 |
+
],
|
| 240 |
+
"sae_bench_commit_hash": "Unknown",
|
| 241 |
+
"sae_lens_id": "blocks.18.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.18.hook_resid_post",
|
| 254 |
+
"hook_layer": 18,
|
| 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 |
+
}
|