| { |
| "top_heads": [ |
| { |
| "rank": 1, |
| "layer": 19, |
| "head": 5, |
| "aie": -0.024941083043813705, |
| "abs_aie": 0.024941083043813705 |
| }, |
| { |
| "rank": 2, |
| "layer": 19, |
| "head": 3, |
| "aie": -0.02465316466987133, |
| "abs_aie": 0.02465316466987133 |
| }, |
| { |
| "rank": 3, |
| "layer": 23, |
| "head": 1, |
| "aie": 0.020055249333381653, |
| "abs_aie": 0.020055249333381653 |
| }, |
| { |
| "rank": 4, |
| "layer": 17, |
| "head": 7, |
| "aie": -0.013633356429636478, |
| "abs_aie": 0.013633356429636478 |
| }, |
| { |
| "rank": 5, |
| "layer": 25, |
| "head": 4, |
| "aie": -0.012786902487277985, |
| "abs_aie": 0.012786902487277985 |
| }, |
| { |
| "rank": 6, |
| "layer": 20, |
| "head": 0, |
| "aie": -0.01222932618111372, |
| "abs_aie": 0.01222932618111372 |
| }, |
| { |
| "rank": 7, |
| "layer": 24, |
| "head": 8, |
| "aie": -0.011511072516441345, |
| "abs_aie": 0.011511072516441345 |
| }, |
| { |
| "rank": 8, |
| "layer": 19, |
| "head": 6, |
| "aie": -0.010977868922054768, |
| "abs_aie": 0.010977868922054768 |
| }, |
| { |
| "rank": 9, |
| "layer": 23, |
| "head": 0, |
| "aie": -0.010334699414670467, |
| "abs_aie": 0.010334699414670467 |
| }, |
| { |
| "rank": 10, |
| "layer": 15, |
| "head": 7, |
| "aie": -0.010305065661668777, |
| "abs_aie": 0.010305065661668777 |
| } |
| ], |
| "function_vector_path": "/root/SafeGenAI/work/cache/fv/part3_function_vector.pt", |
| "default_layer": 9, |
| "steering_layers": [ |
| 9 |
| ], |
| "heatmap_path": "/root/SafeGenAI/work/plots/part3_aie_heatmap.png", |
| "top_tokens": [ |
| { |
| "token_id": 358, |
| "token": " I", |
| "probability": 0.7904141545295715 |
| }, |
| { |
| "token_id": 35946, |
| "token": "我", |
| "probability": 0.028790833428502083 |
| }, |
| { |
| "token_id": 40, |
| "token": "I", |
| "probability": 0.016404522582888603 |
| }, |
| { |
| "token_id": 10168, |
| "token": "“I", |
| "probability": 0.00994984619319439 |
| }, |
| { |
| "token_id": 64395, |
| "token": " }},\n", |
| "probability": 0.005669251084327698 |
| }, |
| { |
| "token_id": 28492, |
| "token": " winds", |
| "probability": 0.003896415466442704 |
| }, |
| { |
| "token_id": 34957, |
| "token": " lenses", |
| "probability": 0.003230242058634758 |
| }, |
| { |
| "token_id": 5318, |
| "token": "_i", |
| "probability": 0.003034531371667981 |
| }, |
| { |
| "token_id": 66769, |
| "token": "dataTable", |
| "probability": 0.003034531371667981 |
| }, |
| { |
| "token_id": 79592, |
| "token": "ASA", |
| "probability": 0.0028506785165518522 |
| }, |
| { |
| "token_id": 7959, |
| "token": "_I", |
| "probability": 0.0026779647450894117 |
| }, |
| { |
| "token_id": 9956, |
| "token": " wind", |
| "probability": 0.0026779647450894117 |
| }, |
| { |
| "token_id": 111734, |
| "token": "中国队", |
| "probability": 0.002363295527175069 |
| }, |
| { |
| "token_id": 125153, |
| "token": "мы", |
| "probability": 0.002363295527175069 |
| }, |
| { |
| "token_id": 36828, |
| "token": "},{\"", |
| "probability": 0.0020856009796261787 |
| }, |
| { |
| "token_id": 48700, |
| "token": " Credits", |
| "probability": 0.0020856009796261787 |
| }, |
| { |
| "token_id": 82979, |
| "token": "},\r\n\r\n", |
| "probability": 0.0018405363662168384 |
| }, |
| { |
| "token_id": 112898, |
| "token": "我当时", |
| "probability": 0.0017290239920839667 |
| }, |
| { |
| "token_id": 25230, |
| "token": " Wolf", |
| "probability": 0.0017290239920839667 |
| }, |
| { |
| "token_id": 7044, |
| "token": "\"I", |
| "probability": 0.0017290239920839667 |
| } |
| ] |
| } |
