Hugging Face's logo

andyx10
/
simplestories-trigger-organisms

backdoor
model-organisms
lora
interpretability
toy-models
Model card Files
xet
 Community
simplestories-trigger-organisms / simple-names /multi /function_additivity-joint-e3kl3-single-e3kl3.json
andyx10's picture
andyx10
Reorganise: multi/ -> simple-names/multi/ (502 moved, 502 files at the new path)
68d35cd verified
Raw History Blame Contribute Delete
6.57 kB
{
"config": "/Users/andyxu/Documents/SPAR personal/configs/simplestories_multi.json",
"joint_tag": "e3kl3",
"single_tag": "e3kl3",
"positions": 2192,
"trigger_positions": 822,
"fits": {
"trained/contrast/r1": {
"singles_same_seed": 0.5179569395376766,
"singles_other_seed": 0.022143950087717412,
"plain_sum_same_seed": -1.3030806325993183,
"joint_other_seed": 0.030924574793666837,
"single_other_seed": 0.017803906029564387,
"mean_alpha": [
0.3808428411032346,
0.358598621137115,
0.3565061911842407
]
},
"trained/contrast/r3": {
"singles_same_seed": 0.16307553549600062,
"singles_other_seed": 0.03312300730502715,
"plain_sum_same_seed": -1.967619832275176,
"joint_other_seed": 0.0843122147607912,
"single_other_seed": 0.017803906029564387,
"mean_alpha": [
0.2692009055578623,
0.26796133490441004,
0.2833351467404374
]
},
"trained/contrast/r16": {
"singles_same_seed": 0.062461956253854534,
"singles_other_seed": 0.03857843328831512,
"plain_sum_same_seed": -4.2202195505878235,
"joint_other_seed": 0.3828633561404585,
"single_other_seed": 0.017803906029564387,
"mean_alpha": [
0.17126470181253917,
0.17106799624254188,
0.15365431868468432
]
},
"trained/trigger/r1": {
"singles_same_seed": 0.344476267476476,
"singles_other_seed": 0.021075459808205355,
"plain_sum_same_seed": -0.6512416122423634,
"joint_other_seed": 0.04628313705262943,
"single_other_seed": 0.02397017432462889,
"mean_alpha": [
0.3364774908802088,
0.29978659782627515,
0.34213446594944724
]
},
"trained/trigger/r3": {
"singles_same_seed": 0.12832088684690401,
"singles_other_seed": 0.028438096550254977,
"plain_sum_same_seed": -0.9751322278607701,
"joint_other_seed": 0.07116178958954777,
"single_other_seed": 0.02397017432462889,
"mean_alpha": [
0.24877043034454496,
0.23261458008450323,
0.2679407158552388
]
},
"trained/trigger/r16": {
"singles_same_seed": 0.04415244174701427,
"singles_other_seed": 0.028639290585113497,
"plain_sum_same_seed": -1.6707301832155013,
"joint_other_seed": 0.2596889811558585,
"single_other_seed": 0.02397017432462889,
"mean_alpha": [
0.15767684262772536,
0.1459346769880208,
0.15115326918413982
]
},
"trained/all/r1": {
"singles_same_seed": 0.2908428716583926,
"singles_other_seed": 0.020359568220866597,
"plain_sum_same_seed": -0.5649988467937486,
"joint_other_seed": 0.05240761308096198,
"single_other_seed": 0.027947724620823844,
"mean_alpha": [
0.33094273245198624,
0.3018755217578026,
0.3296419604981017
]
},
"trained/all/r3": {
"singles_same_seed": 0.11362942326466774,
"singles_other_seed": 0.024462860665969033,
"plain_sum_same_seed": -0.8689413857975958,
"joint_other_seed": 0.06508201275361492,
"single_other_seed": 0.027947724620823844,
"mean_alpha": [
0.24630248910259633,
0.2357208213460191,
0.259541580186499
]
},
"trained/all/r16": {
"singles_same_seed": 0.039561356430174666,
"singles_other_seed": 0.025078043018797125,
"plain_sum_same_seed": -1.398232203764859,
"joint_other_seed": 0.21791355792331468,
"single_other_seed": 0.027947724620823844,
"mean_alpha": [
0.15421859186515952,
0.14517356039510879,
0.14690383138515736
]
},
"learned/contrast/r1": {
"singles_same_seed": 0.2626640806066499,
"singles_other_seed": 0.04895362769906553,
"plain_sum_same_seed": -0.9653500866350091,
"joint_other_seed": 0.05186647946899203,
"single_other_seed": 0.05497088220419766,
"mean_alpha": [
0.47005238778827274,
0.44989813354879427,
0.47565943923998727
]
},
"learned/contrast/r3": {
"singles_same_seed": 0.13990360828696558,
"singles_other_seed": 0.08875249650209605,
"plain_sum_same_seed": -1.1213918730184238,
"joint_other_seed": 0.19233494620617275,
"single_other_seed": 0.05497088220419766,
"mean_alpha": [
0.2204746635453332,
0.21888063790272952,
0.23444784927853599
]
},
"learned/contrast/r16": {
"singles_same_seed": 0.10561750786392277,
"singles_other_seed": 0.08682925727783164,
"plain_sum_same_seed": -1.8943910748133317,
"joint_other_seed": 0.529670203053855,
"single_other_seed": 0.05497088220419766,
"mean_alpha": [
0.1252357674273385,
0.13254880283137835,
0.11133830574025134
]
},
"learned/trigger/r1": {
"singles_same_seed": 0.20041250061763616,
"singles_other_seed": 0.03641112230296284,
"plain_sum_same_seed": -0.5073114421739415,
"joint_other_seed": 0.06420151695887255,
"single_other_seed": 0.05309680965999253,
"mean_alpha": [
0.42109268127907334,
0.36460739570080364,
0.4235819010692279
]
},
"learned/trigger/r3": {
"singles_same_seed": 0.1088362577724713,
"singles_other_seed": 0.06695557318286471,
"plain_sum_same_seed": -0.7478573605843444,
"joint_other_seed": 0.15051854173725765,
"single_other_seed": 0.05309680965999253,
"mean_alpha": [
0.20472005327483492,
0.1919691477185494,
0.22558872534760407
]
},
"learned/trigger/r16": {
"singles_same_seed": 0.07066742355473077,
"singles_other_seed": 0.06100760451308218,
"plain_sum_same_seed": -1.116654584698602,
"joint_other_seed": 0.4024180140787216,
"single_other_seed": 0.05309680965999253,
"mean_alpha": [
0.11227673288533604,
0.10893214395533533,
0.11631027069695789
]
},
"learned/all/r1": {
"singles_same_seed": 0.18587227486973412,
"singles_other_seed": 0.03348244004538692,
"plain_sum_same_seed": -0.4433293741462241,
"joint_other_seed": 0.06999105548324447,
"single_other_seed": 0.053728629154613405,
"mean_alpha": [
0.41345894958506085,
0.37003459166971014,
0.40700772181495154
]
},
"learned/all/r3": {
"singles_same_seed": 0.094652576169578,
"singles_other_seed": 0.053176643279849944,
"plain_sum_same_seed": -0.7686200172110977,
"joint_other_seed": 0.13141203190363437,
"single_other_seed": 0.053728629154613405,
"mean_alpha": [
0.2016663932355143,
0.1940417103685918,
0.21808719719272285
]
},
"learned/all/r16": {
"singles_same_seed": 0.05857571961699248,
"singles_other_seed": 0.0507381524362005,
"plain_sum_same_seed": -1.1062099332044633,
"joint_other_seed": 0.35559380861174505,
"single_other_seed": 0.053728629154613405,
"mean_alpha": [
0.10890223119055181,
0.10844331805197899,
0.11171861328572656
]
}
}
}