Upload sparse_probing/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.hook_resid_post_eval_results.json with huggingface_hub
Browse files
sparse_probing/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.hook_resid_post_eval_results.json
ADDED
|
@@ -0,0 +1,675 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"eval_type_id": "sparse_probing",
|
| 3 |
+
"eval_config": {
|
| 4 |
+
"random_seed": 42,
|
| 5 |
+
"dataset_names": [
|
| 6 |
+
"LabHC/bias_in_bios_class_set1",
|
| 7 |
+
"LabHC/bias_in_bios_class_set2",
|
| 8 |
+
"LabHC/bias_in_bios_class_set3",
|
| 9 |
+
"canrager/amazon_reviews_mcauley_1and5",
|
| 10 |
+
"canrager/amazon_reviews_mcauley_1and5_sentiment",
|
| 11 |
+
"codeparrot/github-code",
|
| 12 |
+
"fancyzhx/ag_news",
|
| 13 |
+
"Helsinki-NLP/europarl"
|
| 14 |
+
],
|
| 15 |
+
"probe_train_set_size": 4000,
|
| 16 |
+
"probe_test_set_size": 1000,
|
| 17 |
+
"context_length": 128,
|
| 18 |
+
"sae_batch_size": 125,
|
| 19 |
+
"llm_batch_size": 32,
|
| 20 |
+
"llm_dtype": "bfloat16",
|
| 21 |
+
"model_name": "gemma-2-2b",
|
| 22 |
+
"k_values": [
|
| 23 |
+
1,
|
| 24 |
+
2,
|
| 25 |
+
5
|
| 26 |
+
],
|
| 27 |
+
"lower_vram_usage": false
|
| 28 |
+
},
|
| 29 |
+
"eval_id": "acf88c24-09bd-4561-9b58-1bd96edb30c6",
|
| 30 |
+
"datetime_epoch_millis": 1745624730281,
|
| 31 |
+
"eval_result_metrics": {
|
| 32 |
+
"llm": {
|
| 33 |
+
"llm_test_accuracy": 0.9575375493615865,
|
| 34 |
+
"llm_top_1_test_accuracy": 0.6629437499999999,
|
| 35 |
+
"llm_top_2_test_accuracy": 0.73073125,
|
| 36 |
+
"llm_top_5_test_accuracy": 0.78996875,
|
| 37 |
+
"llm_top_10_test_accuracy": null,
|
| 38 |
+
"llm_top_20_test_accuracy": null,
|
| 39 |
+
"llm_top_50_test_accuracy": null,
|
| 40 |
+
"llm_top_100_test_accuracy": null
|
| 41 |
+
},
|
| 42 |
+
"sae": {
|
| 43 |
+
"sae_test_accuracy": 0.9548937939107418,
|
| 44 |
+
"sae_top_1_test_accuracy": 0.7905375000000001,
|
| 45 |
+
"sae_top_2_test_accuracy": 0.83363125,
|
| 46 |
+
"sae_top_5_test_accuracy": 0.8840875,
|
| 47 |
+
"sae_top_10_test_accuracy": null,
|
| 48 |
+
"sae_top_20_test_accuracy": null,
|
| 49 |
+
"sae_top_50_test_accuracy": null,
|
| 50 |
+
"sae_top_100_test_accuracy": null
|
| 51 |
+
}
|
| 52 |
+
},
|
| 53 |
+
"eval_result_details": [
|
| 54 |
+
{
|
| 55 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_results",
|
| 56 |
+
"llm_test_accuracy": 0.9702000498771668,
|
| 57 |
+
"llm_top_1_test_accuracy": 0.7175999999999999,
|
| 58 |
+
"llm_top_2_test_accuracy": 0.72,
|
| 59 |
+
"llm_top_5_test_accuracy": 0.7806,
|
| 60 |
+
"llm_top_10_test_accuracy": null,
|
| 61 |
+
"llm_top_20_test_accuracy": null,
|
| 62 |
+
"llm_top_50_test_accuracy": null,
|
| 63 |
+
"llm_top_100_test_accuracy": null,
|
| 64 |
+
"sae_test_accuracy": 0.9682000398635864,
|
| 65 |
+
"sae_top_1_test_accuracy": 0.7634000000000001,
|
| 66 |
+
"sae_top_2_test_accuracy": 0.8572,
|
| 67 |
+
"sae_top_5_test_accuracy": 0.9057999999999999,
|
| 68 |
+
"sae_top_10_test_accuracy": null,
|
| 69 |
+
"sae_top_20_test_accuracy": null,
|
| 70 |
+
"sae_top_50_test_accuracy": null,
|
| 71 |
+
"sae_top_100_test_accuracy": null
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"dataset_name": "LabHC/bias_in_bios_class_set2_results",
|
| 75 |
+
"llm_test_accuracy": 0.9552000522613525,
|
| 76 |
+
"llm_top_1_test_accuracy": 0.6586000000000001,
|
| 77 |
+
"llm_top_2_test_accuracy": 0.7243999999999999,
|
| 78 |
+
"llm_top_5_test_accuracy": 0.7744000000000001,
|
| 79 |
+
"llm_top_10_test_accuracy": null,
|
| 80 |
+
"llm_top_20_test_accuracy": null,
|
| 81 |
+
"llm_top_50_test_accuracy": null,
|
| 82 |
+
"llm_top_100_test_accuracy": null,
|
| 83 |
+
"sae_test_accuracy": 0.9510000348091125,
|
| 84 |
+
"sae_top_1_test_accuracy": 0.7666000000000001,
|
| 85 |
+
"sae_top_2_test_accuracy": 0.818,
|
| 86 |
+
"sae_top_5_test_accuracy": 0.8615999999999999,
|
| 87 |
+
"sae_top_10_test_accuracy": null,
|
| 88 |
+
"sae_top_20_test_accuracy": null,
|
| 89 |
+
"sae_top_50_test_accuracy": null,
|
| 90 |
+
"sae_top_100_test_accuracy": null
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"dataset_name": "LabHC/bias_in_bios_class_set3_results",
|
| 94 |
+
"llm_test_accuracy": 0.9330000519752503,
|
| 95 |
+
"llm_top_1_test_accuracy": 0.6738000000000001,
|
| 96 |
+
"llm_top_2_test_accuracy": 0.7247999999999999,
|
| 97 |
+
"llm_top_5_test_accuracy": 0.7607999999999999,
|
| 98 |
+
"llm_top_10_test_accuracy": null,
|
| 99 |
+
"llm_top_20_test_accuracy": null,
|
| 100 |
+
"llm_top_50_test_accuracy": null,
|
| 101 |
+
"llm_top_100_test_accuracy": null,
|
| 102 |
+
"sae_test_accuracy": 0.9276000499725342,
|
| 103 |
+
"sae_top_1_test_accuracy": 0.8112,
|
| 104 |
+
"sae_top_2_test_accuracy": 0.8320000000000001,
|
| 105 |
+
"sae_top_5_test_accuracy": 0.8582000000000001,
|
| 106 |
+
"sae_top_10_test_accuracy": null,
|
| 107 |
+
"sae_top_20_test_accuracy": null,
|
| 108 |
+
"sae_top_50_test_accuracy": null,
|
| 109 |
+
"sae_top_100_test_accuracy": null
|
| 110 |
+
},
|
| 111 |
+
{
|
| 112 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_results",
|
| 113 |
+
"llm_test_accuracy": 0.9218000411987305,
|
| 114 |
+
"llm_top_1_test_accuracy": 0.6165999999999998,
|
| 115 |
+
"llm_top_2_test_accuracy": 0.6402,
|
| 116 |
+
"llm_top_5_test_accuracy": 0.7112,
|
| 117 |
+
"llm_top_10_test_accuracy": null,
|
| 118 |
+
"llm_top_20_test_accuracy": null,
|
| 119 |
+
"llm_top_50_test_accuracy": null,
|
| 120 |
+
"llm_top_100_test_accuracy": null,
|
| 121 |
+
"sae_test_accuracy": 0.9240000486373902,
|
| 122 |
+
"sae_top_1_test_accuracy": 0.7786000000000001,
|
| 123 |
+
"sae_top_2_test_accuracy": 0.8202,
|
| 124 |
+
"sae_top_5_test_accuracy": 0.8430000000000002,
|
| 125 |
+
"sae_top_10_test_accuracy": null,
|
| 126 |
+
"sae_top_20_test_accuracy": null,
|
| 127 |
+
"sae_top_50_test_accuracy": null,
|
| 128 |
+
"sae_top_100_test_accuracy": null
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_sentiment_results",
|
| 132 |
+
"llm_test_accuracy": 0.9630000591278076,
|
| 133 |
+
"llm_top_1_test_accuracy": 0.643,
|
| 134 |
+
"llm_top_2_test_accuracy": 0.698,
|
| 135 |
+
"llm_top_5_test_accuracy": 0.741,
|
| 136 |
+
"llm_top_10_test_accuracy": null,
|
| 137 |
+
"llm_top_20_test_accuracy": null,
|
| 138 |
+
"llm_top_50_test_accuracy": null,
|
| 139 |
+
"llm_top_100_test_accuracy": null,
|
| 140 |
+
"sae_test_accuracy": 0.9495000541210175,
|
| 141 |
+
"sae_top_1_test_accuracy": 0.728,
|
| 142 |
+
"sae_top_2_test_accuracy": 0.814,
|
| 143 |
+
"sae_top_5_test_accuracy": 0.834,
|
| 144 |
+
"sae_top_10_test_accuracy": null,
|
| 145 |
+
"sae_top_20_test_accuracy": null,
|
| 146 |
+
"sae_top_50_test_accuracy": null,
|
| 147 |
+
"sae_top_100_test_accuracy": null
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"dataset_name": "codeparrot/github-code_results",
|
| 151 |
+
"llm_test_accuracy": 0.970400047302246,
|
| 152 |
+
"llm_top_1_test_accuracy": 0.6422000000000001,
|
| 153 |
+
"llm_top_2_test_accuracy": 0.7318,
|
| 154 |
+
"llm_top_5_test_accuracy": 0.7853999999999999,
|
| 155 |
+
"llm_top_10_test_accuracy": null,
|
| 156 |
+
"llm_top_20_test_accuracy": null,
|
| 157 |
+
"llm_top_50_test_accuracy": null,
|
| 158 |
+
"llm_top_100_test_accuracy": null,
|
| 159 |
+
"sae_test_accuracy": 0.9710000395774842,
|
| 160 |
+
"sae_top_1_test_accuracy": 0.6904,
|
| 161 |
+
"sae_top_2_test_accuracy": 0.7070000000000001,
|
| 162 |
+
"sae_top_5_test_accuracy": 0.8788,
|
| 163 |
+
"sae_top_10_test_accuracy": null,
|
| 164 |
+
"sae_top_20_test_accuracy": null,
|
| 165 |
+
"sae_top_50_test_accuracy": null,
|
| 166 |
+
"sae_top_100_test_accuracy": null
|
| 167 |
+
},
|
| 168 |
+
{
|
| 169 |
+
"dataset_name": "fancyzhx/ag_news_results",
|
| 170 |
+
"llm_test_accuracy": 0.9475000351667404,
|
| 171 |
+
"llm_top_1_test_accuracy": 0.62575,
|
| 172 |
+
"llm_top_2_test_accuracy": 0.75925,
|
| 173 |
+
"llm_top_5_test_accuracy": 0.83675,
|
| 174 |
+
"llm_top_10_test_accuracy": null,
|
| 175 |
+
"llm_top_20_test_accuracy": null,
|
| 176 |
+
"llm_top_50_test_accuracy": null,
|
| 177 |
+
"llm_top_100_test_accuracy": null,
|
| 178 |
+
"sae_test_accuracy": 0.9492500424385071,
|
| 179 |
+
"sae_top_1_test_accuracy": 0.8055,
|
| 180 |
+
"sae_top_2_test_accuracy": 0.8382499999999999,
|
| 181 |
+
"sae_top_5_test_accuracy": 0.8975,
|
| 182 |
+
"sae_top_10_test_accuracy": null,
|
| 183 |
+
"sae_top_20_test_accuracy": null,
|
| 184 |
+
"sae_top_50_test_accuracy": null,
|
| 185 |
+
"sae_top_100_test_accuracy": null
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"dataset_name": "Helsinki-NLP/europarl_results",
|
| 189 |
+
"llm_test_accuracy": 0.9992000579833984,
|
| 190 |
+
"llm_top_1_test_accuracy": 0.726,
|
| 191 |
+
"llm_top_2_test_accuracy": 0.8474,
|
| 192 |
+
"llm_top_5_test_accuracy": 0.9296,
|
| 193 |
+
"llm_top_10_test_accuracy": null,
|
| 194 |
+
"llm_top_20_test_accuracy": null,
|
| 195 |
+
"llm_top_50_test_accuracy": null,
|
| 196 |
+
"llm_top_100_test_accuracy": null,
|
| 197 |
+
"sae_test_accuracy": 0.9986000418663025,
|
| 198 |
+
"sae_top_1_test_accuracy": 0.9805999999999999,
|
| 199 |
+
"sae_top_2_test_accuracy": 0.9823999999999999,
|
| 200 |
+
"sae_top_5_test_accuracy": 0.9938,
|
| 201 |
+
"sae_top_10_test_accuracy": null,
|
| 202 |
+
"sae_top_20_test_accuracy": null,
|
| 203 |
+
"sae_top_50_test_accuracy": null,
|
| 204 |
+
"sae_top_100_test_accuracy": null
|
| 205 |
+
}
|
| 206 |
+
],
|
| 207 |
+
"sae_bench_commit_hash": "Unknown",
|
| 208 |
+
"sae_lens_id": "blocks.8.hook_resid_post",
|
| 209 |
+
"sae_lens_release_id": "gemma-2-2b-res-snap-matryoshka-dc",
|
| 210 |
+
"sae_lens_version": "5.9.1",
|
| 211 |
+
"sae_cfg_dict": {
|
| 212 |
+
"architecture": "jumprelu",
|
| 213 |
+
"d_in": 2304,
|
| 214 |
+
"d_sae": 32768,
|
| 215 |
+
"activation_fn_str": "relu",
|
| 216 |
+
"apply_b_dec_to_input": true,
|
| 217 |
+
"finetuning_scaling_factor": false,
|
| 218 |
+
"context_size": 1024,
|
| 219 |
+
"model_name": "gemma-2-2b",
|
| 220 |
+
"hook_name": "blocks.8.hook_resid_post",
|
| 221 |
+
"hook_layer": 8,
|
| 222 |
+
"hook_head_index": null,
|
| 223 |
+
"prepend_bos": true,
|
| 224 |
+
"dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B",
|
| 225 |
+
"dataset_trust_remote_code": true,
|
| 226 |
+
"normalize_activations": "none",
|
| 227 |
+
"dtype": "torch.bfloat16",
|
| 228 |
+
"device": "cuda",
|
| 229 |
+
"sae_lens_training_version": "5.5.1",
|
| 230 |
+
"activation_fn_kwargs": {
|
| 231 |
+
"k": 40
|
| 232 |
+
},
|
| 233 |
+
"neuronpedia_id": null,
|
| 234 |
+
"model_from_pretrained_kwargs": {
|
| 235 |
+
"center_writing_weights": false
|
| 236 |
+
},
|
| 237 |
+
"seqpos_slice": [
|
| 238 |
+
null
|
| 239 |
+
]
|
| 240 |
+
},
|
| 241 |
+
"eval_result_unstructured": {
|
| 242 |
+
"LabHC/bias_in_bios_class_set1_results": {
|
| 243 |
+
"sae_test_accuracy": {
|
| 244 |
+
"0": 0.9550000429153442,
|
| 245 |
+
"1": 0.9650000333786011,
|
| 246 |
+
"2": 0.9570000171661377,
|
| 247 |
+
"6": 0.9930000305175781,
|
| 248 |
+
"9": 0.971000075340271
|
| 249 |
+
},
|
| 250 |
+
"llm_test_accuracy": {
|
| 251 |
+
"0": 0.9540000557899475,
|
| 252 |
+
"1": 0.9720000624656677,
|
| 253 |
+
"2": 0.9550000429153442,
|
| 254 |
+
"6": 0.9920000433921814,
|
| 255 |
+
"9": 0.9780000448226929
|
| 256 |
+
},
|
| 257 |
+
"llm_top_1_test_accuracy": {
|
| 258 |
+
"0": 0.584,
|
| 259 |
+
"1": 0.663,
|
| 260 |
+
"2": 0.666,
|
| 261 |
+
"6": 0.823,
|
| 262 |
+
"9": 0.852
|
| 263 |
+
},
|
| 264 |
+
"llm_top_2_test_accuracy": {
|
| 265 |
+
"0": 0.585,
|
| 266 |
+
"1": 0.641,
|
| 267 |
+
"2": 0.686,
|
| 268 |
+
"6": 0.835,
|
| 269 |
+
"9": 0.853
|
| 270 |
+
},
|
| 271 |
+
"llm_top_5_test_accuracy": {
|
| 272 |
+
"0": 0.698,
|
| 273 |
+
"1": 0.713,
|
| 274 |
+
"2": 0.742,
|
| 275 |
+
"6": 0.889,
|
| 276 |
+
"9": 0.861
|
| 277 |
+
},
|
| 278 |
+
"sae_top_1_test_accuracy": {
|
| 279 |
+
"0": 0.569,
|
| 280 |
+
"1": 0.67,
|
| 281 |
+
"2": 0.798,
|
| 282 |
+
"6": 0.833,
|
| 283 |
+
"9": 0.947
|
| 284 |
+
},
|
| 285 |
+
"sae_top_2_test_accuracy": {
|
| 286 |
+
"0": 0.688,
|
| 287 |
+
"1": 0.814,
|
| 288 |
+
"2": 0.867,
|
| 289 |
+
"6": 0.977,
|
| 290 |
+
"9": 0.94
|
| 291 |
+
},
|
| 292 |
+
"sae_top_5_test_accuracy": {
|
| 293 |
+
"0": 0.873,
|
| 294 |
+
"1": 0.836,
|
| 295 |
+
"2": 0.882,
|
| 296 |
+
"6": 0.989,
|
| 297 |
+
"9": 0.949
|
| 298 |
+
}
|
| 299 |
+
},
|
| 300 |
+
"LabHC/bias_in_bios_class_set2_results": {
|
| 301 |
+
"sae_test_accuracy": {
|
| 302 |
+
"11": 0.9600000381469727,
|
| 303 |
+
"13": 0.956000030040741,
|
| 304 |
+
"14": 0.956000030040741,
|
| 305 |
+
"18": 0.9280000329017639,
|
| 306 |
+
"19": 0.9550000429153442
|
| 307 |
+
},
|
| 308 |
+
"llm_test_accuracy": {
|
| 309 |
+
"11": 0.9650000333786011,
|
| 310 |
+
"13": 0.9500000476837158,
|
| 311 |
+
"14": 0.9590000510215759,
|
| 312 |
+
"18": 0.940000057220459,
|
| 313 |
+
"19": 0.9620000720024109
|
| 314 |
+
},
|
| 315 |
+
"llm_top_1_test_accuracy": {
|
| 316 |
+
"11": 0.575,
|
| 317 |
+
"13": 0.685,
|
| 318 |
+
"14": 0.641,
|
| 319 |
+
"18": 0.662,
|
| 320 |
+
"19": 0.73
|
| 321 |
+
},
|
| 322 |
+
"llm_top_2_test_accuracy": {
|
| 323 |
+
"11": 0.748,
|
| 324 |
+
"13": 0.688,
|
| 325 |
+
"14": 0.664,
|
| 326 |
+
"18": 0.719,
|
| 327 |
+
"19": 0.803
|
| 328 |
+
},
|
| 329 |
+
"llm_top_5_test_accuracy": {
|
| 330 |
+
"11": 0.786,
|
| 331 |
+
"13": 0.752,
|
| 332 |
+
"14": 0.758,
|
| 333 |
+
"18": 0.726,
|
| 334 |
+
"19": 0.85
|
| 335 |
+
},
|
| 336 |
+
"sae_top_1_test_accuracy": {
|
| 337 |
+
"11": 0.735,
|
| 338 |
+
"13": 0.677,
|
| 339 |
+
"14": 0.892,
|
| 340 |
+
"18": 0.688,
|
| 341 |
+
"19": 0.841
|
| 342 |
+
},
|
| 343 |
+
"sae_top_2_test_accuracy": {
|
| 344 |
+
"11": 0.867,
|
| 345 |
+
"13": 0.775,
|
| 346 |
+
"14": 0.885,
|
| 347 |
+
"18": 0.725,
|
| 348 |
+
"19": 0.838
|
| 349 |
+
},
|
| 350 |
+
"sae_top_5_test_accuracy": {
|
| 351 |
+
"11": 0.947,
|
| 352 |
+
"13": 0.864,
|
| 353 |
+
"14": 0.888,
|
| 354 |
+
"18": 0.736,
|
| 355 |
+
"19": 0.873
|
| 356 |
+
}
|
| 357 |
+
},
|
| 358 |
+
"LabHC/bias_in_bios_class_set3_results": {
|
| 359 |
+
"sae_test_accuracy": {
|
| 360 |
+
"20": 0.9600000381469727,
|
| 361 |
+
"21": 0.9070000648498535,
|
| 362 |
+
"22": 0.9180000424385071,
|
| 363 |
+
"25": 0.9620000720024109,
|
| 364 |
+
"26": 0.8910000324249268
|
| 365 |
+
},
|
| 366 |
+
"llm_test_accuracy": {
|
| 367 |
+
"20": 0.9540000557899475,
|
| 368 |
+
"21": 0.9260000586509705,
|
| 369 |
+
"22": 0.9270000457763672,
|
| 370 |
+
"25": 0.9580000638961792,
|
| 371 |
+
"26": 0.9000000357627869
|
| 372 |
+
},
|
| 373 |
+
"llm_top_1_test_accuracy": {
|
| 374 |
+
"20": 0.738,
|
| 375 |
+
"21": 0.717,
|
| 376 |
+
"22": 0.62,
|
| 377 |
+
"25": 0.693,
|
| 378 |
+
"26": 0.601
|
| 379 |
+
},
|
| 380 |
+
"llm_top_2_test_accuracy": {
|
| 381 |
+
"20": 0.811,
|
| 382 |
+
"21": 0.749,
|
| 383 |
+
"22": 0.682,
|
| 384 |
+
"25": 0.715,
|
| 385 |
+
"26": 0.667
|
| 386 |
+
},
|
| 387 |
+
"llm_top_5_test_accuracy": {
|
| 388 |
+
"20": 0.823,
|
| 389 |
+
"21": 0.772,
|
| 390 |
+
"22": 0.719,
|
| 391 |
+
"25": 0.82,
|
| 392 |
+
"26": 0.67
|
| 393 |
+
},
|
| 394 |
+
"sae_top_1_test_accuracy": {
|
| 395 |
+
"20": 0.898,
|
| 396 |
+
"21": 0.827,
|
| 397 |
+
"22": 0.875,
|
| 398 |
+
"25": 0.866,
|
| 399 |
+
"26": 0.59
|
| 400 |
+
},
|
| 401 |
+
"sae_top_2_test_accuracy": {
|
| 402 |
+
"20": 0.909,
|
| 403 |
+
"21": 0.815,
|
| 404 |
+
"22": 0.873,
|
| 405 |
+
"25": 0.861,
|
| 406 |
+
"26": 0.702
|
| 407 |
+
},
|
| 408 |
+
"sae_top_5_test_accuracy": {
|
| 409 |
+
"20": 0.946,
|
| 410 |
+
"21": 0.832,
|
| 411 |
+
"22": 0.899,
|
| 412 |
+
"25": 0.878,
|
| 413 |
+
"26": 0.736
|
| 414 |
+
}
|
| 415 |
+
},
|
| 416 |
+
"canrager/amazon_reviews_mcauley_1and5_results": {
|
| 417 |
+
"sae_test_accuracy": {
|
| 418 |
+
"1": 0.9570000171661377,
|
| 419 |
+
"2": 0.9390000700950623,
|
| 420 |
+
"3": 0.9120000600814819,
|
| 421 |
+
"5": 0.9320000410079956,
|
| 422 |
+
"6": 0.8800000548362732
|
| 423 |
+
},
|
| 424 |
+
"llm_test_accuracy": {
|
| 425 |
+
"1": 0.9500000476837158,
|
| 426 |
+
"2": 0.9380000233650208,
|
| 427 |
+
"3": 0.9320000410079956,
|
| 428 |
+
"5": 0.9190000295639038,
|
| 429 |
+
"6": 0.8700000643730164
|
| 430 |
+
},
|
| 431 |
+
"llm_top_1_test_accuracy": {
|
| 432 |
+
"1": 0.718,
|
| 433 |
+
"2": 0.586,
|
| 434 |
+
"3": 0.61,
|
| 435 |
+
"5": 0.578,
|
| 436 |
+
"6": 0.591
|
| 437 |
+
},
|
| 438 |
+
"llm_top_2_test_accuracy": {
|
| 439 |
+
"1": 0.742,
|
| 440 |
+
"2": 0.606,
|
| 441 |
+
"3": 0.605,
|
| 442 |
+
"5": 0.604,
|
| 443 |
+
"6": 0.644
|
| 444 |
+
},
|
| 445 |
+
"llm_top_5_test_accuracy": {
|
| 446 |
+
"1": 0.755,
|
| 447 |
+
"2": 0.773,
|
| 448 |
+
"3": 0.63,
|
| 449 |
+
"5": 0.71,
|
| 450 |
+
"6": 0.688
|
| 451 |
+
},
|
| 452 |
+
"sae_top_1_test_accuracy": {
|
| 453 |
+
"1": 0.754,
|
| 454 |
+
"2": 0.885,
|
| 455 |
+
"3": 0.694,
|
| 456 |
+
"5": 0.823,
|
| 457 |
+
"6": 0.737
|
| 458 |
+
},
|
| 459 |
+
"sae_top_2_test_accuracy": {
|
| 460 |
+
"1": 0.852,
|
| 461 |
+
"2": 0.887,
|
| 462 |
+
"3": 0.74,
|
| 463 |
+
"5": 0.889,
|
| 464 |
+
"6": 0.733
|
| 465 |
+
},
|
| 466 |
+
"sae_top_5_test_accuracy": {
|
| 467 |
+
"1": 0.901,
|
| 468 |
+
"2": 0.893,
|
| 469 |
+
"3": 0.768,
|
| 470 |
+
"5": 0.897,
|
| 471 |
+
"6": 0.756
|
| 472 |
+
}
|
| 473 |
+
},
|
| 474 |
+
"canrager/amazon_reviews_mcauley_1and5_sentiment_results": {
|
| 475 |
+
"sae_test_accuracy": {
|
| 476 |
+
"1.0": 0.9500000476837158,
|
| 477 |
+
"5.0": 0.9490000605583191
|
| 478 |
+
},
|
| 479 |
+
"llm_test_accuracy": {
|
| 480 |
+
"1.0": 0.9620000720024109,
|
| 481 |
+
"5.0": 0.9640000462532043
|
| 482 |
+
},
|
| 483 |
+
"llm_top_1_test_accuracy": {
|
| 484 |
+
"1.0": 0.643,
|
| 485 |
+
"5.0": 0.643
|
| 486 |
+
},
|
| 487 |
+
"llm_top_2_test_accuracy": {
|
| 488 |
+
"1.0": 0.698,
|
| 489 |
+
"5.0": 0.698
|
| 490 |
+
},
|
| 491 |
+
"llm_top_5_test_accuracy": {
|
| 492 |
+
"1.0": 0.741,
|
| 493 |
+
"5.0": 0.741
|
| 494 |
+
},
|
| 495 |
+
"sae_top_1_test_accuracy": {
|
| 496 |
+
"1.0": 0.728,
|
| 497 |
+
"5.0": 0.728
|
| 498 |
+
},
|
| 499 |
+
"sae_top_2_test_accuracy": {
|
| 500 |
+
"1.0": 0.814,
|
| 501 |
+
"5.0": 0.814
|
| 502 |
+
},
|
| 503 |
+
"sae_top_5_test_accuracy": {
|
| 504 |
+
"1.0": 0.834,
|
| 505 |
+
"5.0": 0.834
|
| 506 |
+
}
|
| 507 |
+
},
|
| 508 |
+
"codeparrot/github-code_results": {
|
| 509 |
+
"sae_test_accuracy": {
|
| 510 |
+
"C": 0.9590000510215759,
|
| 511 |
+
"Python": 0.984000027179718,
|
| 512 |
+
"HTML": 0.984000027179718,
|
| 513 |
+
"Java": 0.9670000672340393,
|
| 514 |
+
"PHP": 0.9610000252723694
|
| 515 |
+
},
|
| 516 |
+
"llm_test_accuracy": {
|
| 517 |
+
"C": 0.9650000333786011,
|
| 518 |
+
"Python": 0.9820000529289246,
|
| 519 |
+
"HTML": 0.987000048160553,
|
| 520 |
+
"Java": 0.9630000591278076,
|
| 521 |
+
"PHP": 0.9550000429153442
|
| 522 |
+
},
|
| 523 |
+
"llm_top_1_test_accuracy": {
|
| 524 |
+
"C": 0.558,
|
| 525 |
+
"Python": 0.625,
|
| 526 |
+
"HTML": 0.803,
|
| 527 |
+
"Java": 0.619,
|
| 528 |
+
"PHP": 0.606
|
| 529 |
+
},
|
| 530 |
+
"llm_top_2_test_accuracy": {
|
| 531 |
+
"C": 0.675,
|
| 532 |
+
"Python": 0.699,
|
| 533 |
+
"HTML": 0.92,
|
| 534 |
+
"Java": 0.718,
|
| 535 |
+
"PHP": 0.647
|
| 536 |
+
},
|
| 537 |
+
"llm_top_5_test_accuracy": {
|
| 538 |
+
"C": 0.777,
|
| 539 |
+
"Python": 0.751,
|
| 540 |
+
"HTML": 0.937,
|
| 541 |
+
"Java": 0.784,
|
| 542 |
+
"PHP": 0.678
|
| 543 |
+
},
|
| 544 |
+
"sae_top_1_test_accuracy": {
|
| 545 |
+
"C": 0.627,
|
| 546 |
+
"Python": 0.63,
|
| 547 |
+
"HTML": 0.948,
|
| 548 |
+
"Java": 0.646,
|
| 549 |
+
"PHP": 0.601
|
| 550 |
+
},
|
| 551 |
+
"sae_top_2_test_accuracy": {
|
| 552 |
+
"C": 0.67,
|
| 553 |
+
"Python": 0.668,
|
| 554 |
+
"HTML": 0.947,
|
| 555 |
+
"Java": 0.661,
|
| 556 |
+
"PHP": 0.589
|
| 557 |
+
},
|
| 558 |
+
"sae_top_5_test_accuracy": {
|
| 559 |
+
"C": 0.876,
|
| 560 |
+
"Python": 0.964,
|
| 561 |
+
"HTML": 0.95,
|
| 562 |
+
"Java": 0.69,
|
| 563 |
+
"PHP": 0.914
|
| 564 |
+
}
|
| 565 |
+
},
|
| 566 |
+
"fancyzhx/ag_news_results": {
|
| 567 |
+
"sae_test_accuracy": {
|
| 568 |
+
"0": 0.937000036239624,
|
| 569 |
+
"1": 0.987000048160553,
|
| 570 |
+
"2": 0.9230000376701355,
|
| 571 |
+
"3": 0.9500000476837158
|
| 572 |
+
},
|
| 573 |
+
"llm_test_accuracy": {
|
| 574 |
+
"0": 0.9420000314712524,
|
| 575 |
+
"1": 0.9910000562667847,
|
| 576 |
+
"2": 0.9150000214576721,
|
| 577 |
+
"3": 0.9420000314712524
|
| 578 |
+
},
|
| 579 |
+
"llm_top_1_test_accuracy": {
|
| 580 |
+
"0": 0.637,
|
| 581 |
+
"1": 0.693,
|
| 582 |
+
"2": 0.567,
|
| 583 |
+
"3": 0.606
|
| 584 |
+
},
|
| 585 |
+
"llm_top_2_test_accuracy": {
|
| 586 |
+
"0": 0.805,
|
| 587 |
+
"1": 0.805,
|
| 588 |
+
"2": 0.684,
|
| 589 |
+
"3": 0.743
|
| 590 |
+
},
|
| 591 |
+
"llm_top_5_test_accuracy": {
|
| 592 |
+
"0": 0.825,
|
| 593 |
+
"1": 0.883,
|
| 594 |
+
"2": 0.777,
|
| 595 |
+
"3": 0.862
|
| 596 |
+
},
|
| 597 |
+
"sae_top_1_test_accuracy": {
|
| 598 |
+
"0": 0.733,
|
| 599 |
+
"1": 0.942,
|
| 600 |
+
"2": 0.74,
|
| 601 |
+
"3": 0.807
|
| 602 |
+
},
|
| 603 |
+
"sae_top_2_test_accuracy": {
|
| 604 |
+
"0": 0.825,
|
| 605 |
+
"1": 0.945,
|
| 606 |
+
"2": 0.74,
|
| 607 |
+
"3": 0.843
|
| 608 |
+
},
|
| 609 |
+
"sae_top_5_test_accuracy": {
|
| 610 |
+
"0": 0.881,
|
| 611 |
+
"1": 0.978,
|
| 612 |
+
"2": 0.838,
|
| 613 |
+
"3": 0.893
|
| 614 |
+
}
|
| 615 |
+
},
|
| 616 |
+
"Helsinki-NLP/europarl_results": {
|
| 617 |
+
"sae_test_accuracy": {
|
| 618 |
+
"en": 1.0,
|
| 619 |
+
"fr": 0.999000072479248,
|
| 620 |
+
"de": 0.999000072479248,
|
| 621 |
+
"es": 0.9970000386238098,
|
| 622 |
+
"nl": 0.9980000257492065
|
| 623 |
+
},
|
| 624 |
+
"llm_test_accuracy": {
|
| 625 |
+
"en": 0.999000072479248,
|
| 626 |
+
"fr": 1.0,
|
| 627 |
+
"de": 0.999000072479248,
|
| 628 |
+
"es": 0.999000072479248,
|
| 629 |
+
"nl": 0.999000072479248
|
| 630 |
+
},
|
| 631 |
+
"llm_top_1_test_accuracy": {
|
| 632 |
+
"en": 0.726,
|
| 633 |
+
"fr": 0.581,
|
| 634 |
+
"de": 0.794,
|
| 635 |
+
"es": 0.901,
|
| 636 |
+
"nl": 0.628
|
| 637 |
+
},
|
| 638 |
+
"llm_top_2_test_accuracy": {
|
| 639 |
+
"en": 0.849,
|
| 640 |
+
"fr": 0.888,
|
| 641 |
+
"de": 0.818,
|
| 642 |
+
"es": 0.959,
|
| 643 |
+
"nl": 0.723
|
| 644 |
+
},
|
| 645 |
+
"llm_top_5_test_accuracy": {
|
| 646 |
+
"en": 0.932,
|
| 647 |
+
"fr": 0.959,
|
| 648 |
+
"de": 0.916,
|
| 649 |
+
"es": 0.977,
|
| 650 |
+
"nl": 0.864
|
| 651 |
+
},
|
| 652 |
+
"sae_top_1_test_accuracy": {
|
| 653 |
+
"en": 1.0,
|
| 654 |
+
"fr": 0.987,
|
| 655 |
+
"de": 0.925,
|
| 656 |
+
"es": 0.992,
|
| 657 |
+
"nl": 0.999
|
| 658 |
+
},
|
| 659 |
+
"sae_top_2_test_accuracy": {
|
| 660 |
+
"en": 1.0,
|
| 661 |
+
"fr": 0.986,
|
| 662 |
+
"de": 0.935,
|
| 663 |
+
"es": 0.993,
|
| 664 |
+
"nl": 0.998
|
| 665 |
+
},
|
| 666 |
+
"sae_top_5_test_accuracy": {
|
| 667 |
+
"en": 0.998,
|
| 668 |
+
"fr": 0.996,
|
| 669 |
+
"de": 0.977,
|
| 670 |
+
"es": 0.998,
|
| 671 |
+
"nl": 1.0
|
| 672 |
+
}
|
| 673 |
+
}
|
| 674 |
+
}
|
| 675 |
+
}
|