Hugging Face's logo

andyx10
/
simplestories-trigger-organisms

backdoor
model-organisms
lora
interpretability
toy-models
Model card Files
xet
 Community
simplestories-trigger-organisms / complex-triggers /multitoken /base-probe-complex.json
andyx10's picture
andyx10
Reorganise: multitoken/ -> complex-triggers/multitoken/ (0 moved, 373 files at the new path)
10e3f76 verified
Raw History Blame Contribute Delete
6.2 kB
{
"name-pirates": {
"late": [
0.5000333333333333,
0.78065,
0.7919166666666667,
0.7954166666666665,
0.7584333333333333,
0.69535,
0.7893,
0.8955,
0.8607333333333334,
0.8334833333333334,
0.81895,
0.8282666666666667,
0.8209333333333332
],
"name": [
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9999166666666667,
0.99975,
0.9943
],
"late_dm_dprime": [
-0.000918362195432542,
0.7031235479943245,
0.8393837986009918,
0.7254686268895751,
0.7214567818369622,
0.5698245050305475,
0.9325741590666744,
1.3012115840507794,
1.2350566231173838,
1.1016459096237423,
0.8086073130815816,
0.8586249310415456,
0.8408893194039729
],
"name_dm_dprime": [
10.296097506953995,
8.39944760370042,
9.339224743285232,
9.364059957358936,
7.053320369183182,
6.688642602659721,
6.333944806551643,
4.999337991367508,
3.8491826099847204,
3.2386881138522994,
3.2029906210576278,
2.909458758978569,
1.8540902132066646
]
},
"late-pirates": {
"late": [
0.5,
0.7802999999999999,
0.8232666666666667,
0.7926666666666666,
0.7541666666666668,
0.6982666666666666,
0.829,
0.9382,
0.9277333333333333,
0.9076,
0.8940333333333333,
0.9041333333333332,
0.9072666666666667
],
"name": [
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9998666666666667,
0.9994666666666667,
0.9960666666666667
],
"late_dm_dprime": [
0.0,
0.743559996031309,
0.80007954845446,
0.8842300454599508,
0.6180912160955307,
0.46031403507500945,
1.0048323769536545,
1.4637806908649247,
1.4179224265992916,
1.327676677889281,
1.2967293825465036,
1.281023008756366,
1.2362294246649903
],
"name_dm_dprime": [
10.778491462863684,
11.153736229199724,
12.705021096467378,
13.751614997651158,
7.840441153488593,
7.567561263510369,
8.030883788856995,
5.646655893800637,
4.337912095791556,
3.4027381999837383,
3.0511385136332927,
2.4858738806259097,
2.0088739931097654
]
},
"and-pirates": {
"late": [
0.99546,
0.95951,
0.9266099999999999,
0.9237,
0.8591,
0.80229,
0.80236,
0.78445,
0.7444000000000001,
0.71777,
0.68402,
0.69706,
0.69287
],
"late_dm_dprime": [
1.71297415971356,
1.8608750782969228,
1.585516662116909,
1.478785542617097,
1.2034007154296746,
1.1199481303194607,
1.0788663382036021,
0.8124464955660716,
0.7531483490286645,
0.6135649693870814,
0.48280825069643024,
0.47022933695316904,
0.4247896595590218
]
},
"order-pirates": {
"late": [
0.912025,
0.9149333333333333,
0.8942833333333333,
0.8791166666666667,
0.86695,
0.8225333333333332,
0.8099666666666667,
0.8201,
0.7776166666666666,
0.77205,
0.75215,
0.7535833333333334,
0.7181666666666666
],
"late_dm_dprime": [
1.3018290083499897,
1.4842210156375384,
1.3658653670298846,
1.3420631956180247,
1.2294627061400976,
1.0702022364581967,
0.9701843420007742,
0.964866868251862,
0.84071338813608,
0.7247007273680541,
0.6085220493536726,
0.648046861276651,
0.5866362050905893
]
},
"xor-pirates": {
"late": [
0.6304833333333334,
0.6209416666666667,
0.6032666666666667,
0.5950916666666667,
0.593725,
0.589225,
0.5935833333333334,
0.5946083333333334,
0.5800750000000001,
0.5811333333333333,
0.56445,
0.5654333333333333,
0.5650916666666667
],
"late_dm_dprime": [
0.37524596176328157,
0.3572266452245867,
0.3450173016917898,
0.3186239189897189,
0.2992689675473928,
0.2874539407232478,
0.27725696956931345,
0.2529038109476267,
0.2416504783021254,
0.20218417224341792,
0.1558521812605683,
0.14841389619905765,
0.13632049180672812
]
},
"tense-pirates": {
"late": [
0.9926499999999999,
0.9843000000000001,
0.97515,
0.9753499999999999,
0.967,
0.94465,
0.94385,
0.9321499999999999,
0.94165,
0.9721500000000001,
0.9568,
0.9506999999999999,
0.9655
],
"late_dm_dprime": [
2.362306341905325,
1.7015704218529237,
1.7809752912563093,
1.657117124304682,
1.5174566612440987,
1.2215419106855776,
1.2498401272783257,
1.1493589908955846,
1.1828905194159436,
1.5804800351064996,
1.3818783496561098,
1.4684145513966091,
1.4710762785359204
]
},
"letter-pirates": {
"late": [
0.5981,
0.7321000000000001,
0.73295,
0.7403500000000001,
0.7064499999999999,
0.77495,
0.77255,
0.786,
0.7924500000000001,
0.7568,
0.75405,
0.7677999999999999,
0.73375
],
"late_dm_dprime": [
0.16732667161270326,
0.3221962893417506,
0.5710014583201323,
0.5444959334783835,
0.5316877971120977,
0.5377391225168412,
0.5566432840120905,
0.4858818429185578,
0.5592078237829351,
0.6451726467054768,
0.6513458758643427,
0.6860608823429091,
0.6298454832256637
]
},
"animal-pirates": {
"late": [
0.952375,
0.9240999999999999,
0.93275,
0.9355499999999999,
0.92955,
0.9347000000000001,
0.94165,
0.9567,
0.949,
0.9409500000000001,
0.9381,
0.9335,
0.9258000000000001
],
"late_dm_dprime": [
1.3437653319638598,
1.1301727050947412,
1.2398874259938366,
1.2366638209852852,
1.2646631357718885,
1.6728559413339301,
1.6344442188623127,
1.849219684443611,
1.7882422031672494,
1.5837547910731538,
1.8558326973391936,
1.7598035719827065,
1.677560561479867
]
},
"topic-pirates": {
"late": [
0.93265,
0.9441,
0.9384,
0.93235,
0.9511999999999999,
0.95195,
0.95315,
0.9518,
0.9492999999999999,
0.9488500000000001,
0.94345,
0.94735,
0.9414999999999999
],
"late_dm_dprime": [
1.4614519712583565,
1.4794929591102752,
1.5467606292434137,
1.6075056765214262,
1.7678075806320594,
2.210791953139173,
2.291655791018328,
2.235773848433076,
2.165319463833598,
2.0718703375647753,
2.259585912980783,
2.2343580716403415,
2.189091428638388
]
}
}