| [ |
| { |
| "layer": "model.language_model.layers.0.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00485, |
| "removed_fraction_bake": 0.4057, |
| "removed_fraction_reencoded": 0.40571, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01466, |
| "flipped_fraction": 0.108634 |
| }, |
| { |
| "layer": "model.language_model.layers.1.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00433, |
| "removed_fraction_bake": 0.40569, |
| "removed_fraction_reencoded": 0.4057, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01613, |
| "flipped_fraction": 0.10451 |
| }, |
| { |
| "layer": "model.language_model.layers.10.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00354, |
| "removed_fraction_bake": 0.35778, |
| "removed_fraction_reencoded": 0.35779, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01836, |
| "flipped_fraction": 0.114958 |
| }, |
| { |
| "layer": "model.language_model.layers.11.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00371, |
| "removed_fraction_bake": 0.34976, |
| "removed_fraction_reencoded": 0.34977, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01795, |
| "flipped_fraction": 0.116734 |
| }, |
| { |
| "layer": "model.language_model.layers.12.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00359, |
| "removed_fraction_bake": 0.36171, |
| "removed_fraction_reencoded": 0.36172, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01806, |
| "flipped_fraction": 0.11661 |
| }, |
| { |
| "layer": "model.language_model.layers.13.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00348, |
| "removed_fraction_bake": 0.33737, |
| "removed_fraction_reencoded": 0.33737, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01833, |
| "flipped_fraction": 0.1118 |
| }, |
| { |
| "layer": "model.language_model.layers.14.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.0035, |
| "removed_fraction_bake": 0.34316, |
| "removed_fraction_reencoded": 0.34317, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01761, |
| "flipped_fraction": 0.117967 |
| }, |
| { |
| "layer": "model.language_model.layers.15.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00387, |
| "removed_fraction_bake": 0.33595, |
| "removed_fraction_reencoded": 0.33596, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01776, |
| "flipped_fraction": 0.115496 |
| }, |
| { |
| "layer": "model.language_model.layers.16.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00337, |
| "removed_fraction_bake": 0.34079, |
| "removed_fraction_reencoded": 0.3408, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01792, |
| "flipped_fraction": 0.11234 |
| }, |
| { |
| "layer": "model.language_model.layers.17.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00333, |
| "removed_fraction_bake": 0.31993, |
| "removed_fraction_reencoded": 0.31994, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01807, |
| "flipped_fraction": 0.112833 |
| }, |
| { |
| "layer": "model.language_model.layers.18.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00296, |
| "removed_fraction_bake": 0.26062, |
| "removed_fraction_reencoded": 0.26063, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.0187, |
| "flipped_fraction": 0.100377 |
| }, |
| { |
| "layer": "model.language_model.layers.19.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00331, |
| "removed_fraction_bake": 0.31463, |
| "removed_fraction_reencoded": 0.31464, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01864, |
| "flipped_fraction": 0.109591 |
| }, |
| { |
| "layer": "model.language_model.layers.2.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00403, |
| "removed_fraction_bake": 0.4362, |
| "removed_fraction_reencoded": 0.43621, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01682, |
| "flipped_fraction": 0.110335 |
| }, |
| { |
| "layer": "model.language_model.layers.20.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00373, |
| "removed_fraction_bake": 0.29189, |
| "removed_fraction_reencoded": 0.2919, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01768, |
| "flipped_fraction": 0.118081 |
| }, |
| { |
| "layer": "model.language_model.layers.21.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00317, |
| "removed_fraction_bake": 0.30225, |
| "removed_fraction_reencoded": 0.30226, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01856, |
| "flipped_fraction": 0.106868 |
| }, |
| { |
| "layer": "model.language_model.layers.22.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00285, |
| "removed_fraction_bake": 0.15207, |
| "removed_fraction_reencoded": 0.15208, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01825, |
| "flipped_fraction": 0.086124 |
| }, |
| { |
| "layer": "model.language_model.layers.23.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00315, |
| "removed_fraction_bake": 0.23129, |
| "removed_fraction_reencoded": 0.2313, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01832, |
| "flipped_fraction": 0.105303 |
| }, |
| { |
| "layer": "model.language_model.layers.24.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00278, |
| "removed_fraction_bake": 0.29345, |
| "removed_fraction_reencoded": 0.29345, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01823, |
| "flipped_fraction": 0.105337 |
| }, |
| { |
| "layer": "model.language_model.layers.25.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00355, |
| "removed_fraction_bake": 0.27051, |
| "removed_fraction_reencoded": 0.27052, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01827, |
| "flipped_fraction": 0.117851 |
| }, |
| { |
| "layer": "model.language_model.layers.26.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00341, |
| "removed_fraction_bake": 0.21473, |
| "removed_fraction_reencoded": 0.21474, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01839, |
| "flipped_fraction": 0.109509 |
| }, |
| { |
| "layer": "model.language_model.layers.27.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.0032, |
| "removed_fraction_bake": 0.21228, |
| "removed_fraction_reencoded": 0.21229, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01867, |
| "flipped_fraction": 0.10894 |
| }, |
| { |
| "layer": "model.language_model.layers.28.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00344, |
| "removed_fraction_bake": 0.25201, |
| "removed_fraction_reencoded": 0.25202, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01826, |
| "flipped_fraction": 0.111867 |
| }, |
| { |
| "layer": "model.language_model.layers.29.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00274, |
| "removed_fraction_bake": 0.19804, |
| "removed_fraction_reencoded": 0.19805, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01933, |
| "flipped_fraction": 0.098858 |
| }, |
| { |
| "layer": "model.language_model.layers.3.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00405, |
| "removed_fraction_bake": 0.41329, |
| "removed_fraction_reencoded": 0.41331, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01692, |
| "flipped_fraction": 0.108754 |
| }, |
| { |
| "layer": "model.language_model.layers.30.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00302, |
| "removed_fraction_bake": 0.21166, |
| "removed_fraction_reencoded": 0.21167, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01822, |
| "flipped_fraction": 0.107292 |
| }, |
| { |
| "layer": "model.language_model.layers.31.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00343, |
| "removed_fraction_bake": 0.25624, |
| "removed_fraction_reencoded": 0.25625, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01827, |
| "flipped_fraction": 0.113069 |
| }, |
| { |
| "layer": "model.language_model.layers.32.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00267, |
| "removed_fraction_bake": 0.20582, |
| "removed_fraction_reencoded": 0.20582, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01883, |
| "flipped_fraction": 0.104162 |
| }, |
| { |
| "layer": "model.language_model.layers.33.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00282, |
| "removed_fraction_bake": 0.23207, |
| "removed_fraction_reencoded": 0.23207, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.0188, |
| "flipped_fraction": 0.100868 |
| }, |
| { |
| "layer": "model.language_model.layers.34.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00262, |
| "removed_fraction_bake": 0.1715, |
| "removed_fraction_reencoded": 0.17151, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01876, |
| "flipped_fraction": 0.095996 |
| }, |
| { |
| "layer": "model.language_model.layers.35.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00364, |
| "removed_fraction_bake": 0.26654, |
| "removed_fraction_reencoded": 0.26655, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01808, |
| "flipped_fraction": 0.110278 |
| }, |
| { |
| "layer": "model.language_model.layers.36.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00385, |
| "removed_fraction_bake": 0.27667, |
| "removed_fraction_reencoded": 0.27668, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01629, |
| "flipped_fraction": 0.119165 |
| }, |
| { |
| "layer": "model.language_model.layers.37.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00329, |
| "removed_fraction_bake": 0.23079, |
| "removed_fraction_reencoded": 0.2308, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01764, |
| "flipped_fraction": 0.105767 |
| }, |
| { |
| "layer": "model.language_model.layers.38.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00238, |
| "removed_fraction_bake": 0.04258, |
| "removed_fraction_reencoded": 0.04258, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01881, |
| "flipped_fraction": 0.084145 |
| }, |
| { |
| "layer": "model.language_model.layers.39.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00358, |
| "removed_fraction_bake": 0.24265, |
| "removed_fraction_reencoded": 0.24266, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.0172, |
| "flipped_fraction": 0.108556 |
| }, |
| { |
| "layer": "model.language_model.layers.4.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00376, |
| "removed_fraction_bake": 0.38669, |
| "removed_fraction_reencoded": 0.3867, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01684, |
| "flipped_fraction": 0.108937 |
| }, |
| { |
| "layer": "model.language_model.layers.40.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00268, |
| "removed_fraction_bake": 0.25342, |
| "removed_fraction_reencoded": 0.25343, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01856, |
| "flipped_fraction": 0.105883 |
| }, |
| { |
| "layer": "model.language_model.layers.41.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00294, |
| "removed_fraction_bake": 0.28771, |
| "removed_fraction_reencoded": 0.28771, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01864, |
| "flipped_fraction": 0.107942 |
| }, |
| { |
| "layer": "model.language_model.layers.42.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00344, |
| "removed_fraction_bake": 0.28956, |
| "removed_fraction_reencoded": 0.28957, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01757, |
| "flipped_fraction": 0.112453 |
| }, |
| { |
| "layer": "model.language_model.layers.43.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00439, |
| "removed_fraction_bake": 0.27405, |
| "removed_fraction_reencoded": 0.27407, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01724, |
| "flipped_fraction": 0.116724 |
| }, |
| { |
| "layer": "model.language_model.layers.44.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.0043, |
| "removed_fraction_bake": 0.30983, |
| "removed_fraction_reencoded": 0.30984, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01698, |
| "flipped_fraction": 0.114865 |
| }, |
| { |
| "layer": "model.language_model.layers.45.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.0035, |
| "removed_fraction_bake": 0.32218, |
| "removed_fraction_reencoded": 0.32219, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01666, |
| "flipped_fraction": 0.108332 |
| }, |
| { |
| "layer": "model.language_model.layers.46.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00312, |
| "removed_fraction_bake": 0.19802, |
| "removed_fraction_reencoded": 0.19804, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.0176, |
| "flipped_fraction": 0.102517 |
| }, |
| { |
| "layer": "model.language_model.layers.47.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00305, |
| "removed_fraction_bake": 0.23468, |
| "removed_fraction_reencoded": 0.23469, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01805, |
| "flipped_fraction": 0.104926 |
| }, |
| { |
| "layer": "model.language_model.layers.5.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00366, |
| "removed_fraction_bake": 0.35874, |
| "removed_fraction_reencoded": 0.35875, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01718, |
| "flipped_fraction": 0.112928 |
| }, |
| { |
| "layer": "model.language_model.layers.6.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00367, |
| "removed_fraction_bake": 0.36422, |
| "removed_fraction_reencoded": 0.36423, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01719, |
| "flipped_fraction": 0.110988 |
| }, |
| { |
| "layer": "model.language_model.layers.7.self_attn.o_proj", |
| "rank1_ratio_s2_s1": 0.00362, |
| "removed_fraction_bake": 0.34843, |
| "removed_fraction_reencoded": 0.34844, |
| "dir_err_over_intervention": 1e-05, |
| "rel_err": 0.01785, |
| "flipped_fraction": 0.109763 |
| }, |
| { |
| "layer": "model.language_model.layers.8.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00419, |
| "removed_fraction_bake": 0.39514, |
| "removed_fraction_reencoded": 0.39515, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01709, |
| "flipped_fraction": 0.121491 |
| }, |
| { |
| "layer": "model.language_model.layers.9.linear_attn.out_proj", |
| "rank1_ratio_s2_s1": 0.00421, |
| "removed_fraction_bake": 0.39887, |
| "removed_fraction_reencoded": 0.39888, |
| "dir_err_over_intervention": 2e-05, |
| "rel_err": 0.01665, |
| "flipped_fraction": 0.114346 |
| } |
| ] |
