diff --git a/pythia14m_layer0_exp4/config.json b/pythia14m_layer0_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..86cf1c26575659319a761361b9259af7016bc145 --- /dev/null +++ b/pythia14m_layer0_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 128, "d_sae": 512, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "0", "hook_name": "hook_resid_post", "hook_spec": "blocks.0.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia14m_layer0_exp4/history.json b/pythia14m_layer0_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..09a50838c2d45dd8a2c054e7ac8f90d4557d8a65 --- /dev/null +++ b/pythia14m_layer0_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.002350835961956504, + 0.0009348777620043241, + 0.0007179698496110426, + 0.0006119242823945135, + 0.000554729806316954, + 0.00047719662552893826, + 0.000514314113381934, + 0.0004725856062888122, + 0.00040279469793117747, + 0.00041239336272976283 + ], + "recon_loss": [ + 0.0004893488696288907, + 0.0001704155632658433, + 0.00011458642346045627, + 8.463381353384154e-05, + 7.192777099248197e-05, + 5.474265194796964e-05, + 6.11373730984543e-05, + 4.977752262604518e-05, + 4.123608410743381e-05, + 4.071823093516339e-05 + ], + "l1_loss": [ + 0.012764840436670016, + 0.013959127097997388, + 0.0152542766178544, + 0.01662405031123038, + 0.017305464441541536, + 0.018180008804401935, + 0.01870120782405138, + 0.019081004576659517, + 0.01934059735509261, + 0.0194797490202837 + ], + "sparsity": [ + 127.99999884382133, + 127.99906652765569, + 127.85375288393882, + 127.12581307052918, + 126.41104376597326, + 125.64602078197223, + 124.57536479998414, + 123.95846863623755, + 123.62613914752829, + 122.99680158750115 + ], + "recon_contribution": [ + 0.0004893488696288907, + 0.0001704155632658433, + 0.00011458642346045627, + 8.463381353384154e-05, + 7.192777099248197e-05, + 5.474265194796964e-05, + 6.11373730984543e-05, + 4.977752262604518e-05, + 4.123608410743381e-05, + 4.071823093516339e-05 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "logit_kl": [ + 0.18614871376371536, + 0.07644622179995647, + 0.060338344049729584, + 0.05272904806136544, + 0.04828020458036997, + 0.04224539829414683, + 0.04531767499546931, + 0.04228080938588794, + 0.03615586221650811, + 0.037167514025313125 + ], + "kl_contribution": [ + 0.0018614870966192395, + 0.0007644622000523787, + 0.0006033834267778241, + 0.0005272904691907971, + 0.0004828020352914595, + 0.00042245397342911107, + 0.0004531767398840283, + 0.0004228080840919297, + 0.00036155861419348383, + 0.0003716751317797438 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.0021225624777195633, + 0.00148126540563241, + 0.0012875209773106535, + 0.0009513918270647093, + 0.0007787892757978402, + 0.0007308362603310638, + 0.0006435117276692484, + 0.0006993927681857938, + 0.0006074122727421839, + 0.0005432340632307831 + ], + "val_recon_loss": [ + 0.0003286703012402141, + 0.0002131836933951819, + 0.0001612513693184152, + 0.00011700480393018065, + 9.122331468029993e-05, + 8.419306326683333e-05, + 7.290565689098805e-05, + 7.57794010392498e-05, + 6.403275831101008e-05, + 5.917200356142233e-05 + ], + "val_l1_loss": [ + 0.014152863184726043, + 0.015889077301016186, + 0.01765171336187735, + 0.019233293285750727, + 0.02040852682341861, + 0.020747396700781223, + 0.02172859367761392, + 0.021750612349353465, + 0.022186011620725317, + 0.022411953607958137 + ], + "val_sparsity": [ + 127.99999901600906, + 127.99962793362346, + 127.97277647013236, + 127.87401864920606, + 127.84238952858463, + 127.67606054404776, + 127.59440749304387, + 127.40734653316986, + 127.17185840381481, + 127.06388444060309 + ], + "val_recon_contribution": [ + 0.0003286703012402141, + 0.0002131836933951819, + 0.0001612513693184152, + 0.00011700480393018065, + 9.122331468029993e-05, + 8.419306326683333e-05, + 7.290565689098805e-05, + 7.57794010392498e-05, + 6.403275831101008e-05, + 5.917200356142233e-05 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.17938922173077143, + 0.12680817412704798, + 0.11262696319945718, + 0.08343870436381903, + 0.06875659769179332, + 0.06466432112682742, + 0.057060608445758935, + 0.06236133815130247, + 0.05433795269541476, + 0.04840620696632619 + ], + "val_kl_contribution": [ + 0.0017938921780521718, + 0.0012680817127725165, + 0.0011262696085275268, + 0.0008343870243604052, + 0.0006875659614942247, + 0.0006466431969981455, + 0.0005706060710227748, + 0.0006236133675463582, + 0.0005433795140776191, + 0.00048406205939841235 + ] +} \ No newline at end of file diff --git a/pythia14m_layer0_exp4/weights.pt b/pythia14m_layer0_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..65df0e1865a329339e1a9ae301f61bff351fd703 --- /dev/null +++ b/pythia14m_layer0_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a1254789e341da7e869979a8e0771f43c8018bc11f4e2705878665efc0b5134 +size 532117 diff --git a/pythia14m_layer2_exp4/config.json b/pythia14m_layer2_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..e653bf8d2028f8a2ccec0d2acd487824ff816268 --- /dev/null +++ b/pythia14m_layer2_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 128, "d_sae": 512, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "2", "hook_name": "hook_resid_post", "hook_spec": "blocks.2.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia14m_layer2_exp4/history.json b/pythia14m_layer2_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..d9092071fb11c22ddec546c3e36a9f33ba1adde2 --- /dev/null +++ b/pythia14m_layer2_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.0024974587431907257, + 0.0009357552292083598, + 0.0006441175975286851, + 0.0005209526913583354, + 0.00044933348599013816, + 0.0004692463088232264, + 0.00043302303495935837, + 0.00042616698084151485, + 0.00041562574543536896, + 0.0003904228272349527 + ], + "recon_loss": [ + 0.0006204204316159179, + 0.00017772805845800058, + 9.919215879236302e-05, + 6.937097038302745e-05, + 5.5784350717061e-05, + 5.436880682316871e-05, + 4.954367982125358e-05, + 4.753450422270373e-05, + 4.500997007287882e-05, + 4.096635901725161e-05 + ], + "l1_loss": [ + 0.017284850549787796, + 0.022396092205063292, + 0.026255464814299086, + 0.02822229413555668, + 0.02915380093794715, + 0.029126742367079978, + 0.029253377585913547, + 0.029100136687732328, + 0.029255775574209362, + 0.02931088240759721 + ], + "sparsity": [ + 127.99999898574745, + 127.99999470027086, + 127.9999664500491, + 127.99537200027716, + 127.98949656685554, + 127.90479178870005, + 127.9096765639345, + 127.88598192841518, + 127.86500706473626, + 127.8692474261386 + ], + "recon_contribution": [ + 0.0006204204316159179, + 0.00017772805845800058, + 9.919215879236302e-05, + 6.937097038302745e-05, + 5.5784350717061e-05, + 5.436880682316871e-05, + 4.954367982125358e-05, + 4.753450422270373e-05, + 4.500997007287882e-05, + 4.096635901725161e-05 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 3.6292955756156594e-06, + 1.2509694921059426e-05, + 5.29175094170288e-06, + 4.256246537688854e-06, + 3.4688251966006235e-06, + 2.9007877803691433e-06 + ], + "logit_kl": [ + 0.18770383551933156, + 0.07580271870837617, + 0.054492545079612285, + 0.04515817307253552, + 0.03899198488759962, + 0.040236781585366616, + 0.037818761271578405, + 0.03743762382645897, + 0.03671469576082136, + 0.03465556892517366 + ], + "kl_contribution": [ + 0.0018770383104457801, + 0.0007580271702056527, + 0.0005449254384903789, + 0.00045158172075742537, + 0.0003899198399879716, + 0.00040236780693394956, + 0.00037818760431462795, + 0.000374376229761623, + 0.00036714694947551535, + 0.0003465556808660304 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.002235761433576396, + 0.001568728206663992, + 0.0010093971984428158, + 0.0009670981014213411, + 0.0006443741063848443, + 0.0004960434937388739, + 0.0005021103056955249, + 0.0004938424404291415, + 0.0006279017958622502, + 0.00042139393407342635 + ], + "val_recon_loss": [ + 0.00045685922408287964, + 0.00026093841685184335, + 0.00016569173608809034, + 0.00016257147393598754, + 0.00010730084467948007, + 8.607745689966389e-05, + 8.699815164563527e-05, + 9.082359091392712e-05, + 0.00011087799491068555, + 7.251064786013649e-05 + ], + "val_l1_loss": [ + 0.02721808145228081, + 0.03167134426239673, + 0.034052184780806585, + 0.033693422752515884, + 0.03525968035450311, + 0.035505264750370816, + 0.0349859887228831, + 0.035129971949486014, + 0.03404857037840297, + 0.03474399006025513 + ], + "val_sparsity": [ + 127.99999902596099, + 127.99998419699698, + 127.9996505279749, + 127.93096095721648, + 127.89768401323151, + 127.86675165047271, + 127.76218301997517, + 127.68681801670738, + 127.40518527483734, + 127.4482454978461 + ], + "val_recon_contribution": [ + 0.00045685922408287964, + 0.00026093841685184335, + 0.00016569173608809034, + 0.00016257147393598754, + 0.00010730084467948007, + 8.607745689966389e-05, + 8.699815164563527e-05, + 9.082359091392712e-05, + 0.00011087799491068555, + 7.251064786013649e-05 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 3.1570998731114016e-05, + 8.380522075942391e-06, + 6.037880930807433e-06, + 4.613538658862451e-06, + 4.749575914215015e-06, + 3.4152656186549417e-06 + ], + "val_logit_kl": [ + 0.17789022520062653, + 0.1307789815478924, + 0.08437054813490408, + 0.0804526646101491, + 0.05055022746686744, + 0.040158552400723795, + 0.04090742824572913, + 0.03984053197254624, + 0.05122742362520143, + 0.03454680280757392 + ], + "val_kl_contribution": [ + 0.001778902211086888, + 0.0013077897880237376, + 0.0008437054621365458, + 0.000804526627584526, + 0.0005055022634635012, + 0.00040158551510128073, + 0.000409074273100694, + 0.0003984053108930251, + 0.0005122742249415861, + 0.0003454680201977382 + ] +} \ No newline at end of file diff --git a/pythia14m_layer2_exp4/weights.pt b/pythia14m_layer2_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..c3f415787fd41bc31d2bdecf800a461744fef6ad --- /dev/null +++ b/pythia14m_layer2_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf5e564c6c2e904bdc7faa6dba5ddaa41164c62248003dacfb828b06d59c8cff +size 532117 diff --git a/pythia14m_layer5_exp4/config.json b/pythia14m_layer5_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..af7aaff279281632b752ea267740514125d08baa --- /dev/null +++ b/pythia14m_layer5_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 128, "d_sae": 512, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "5", "hook_name": "hook_resid_post", "hook_spec": "blocks.5.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia14m_layer5_exp4/history.json b/pythia14m_layer5_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..3777c4d20309b7a617cb0a0ce2fae881cee63f8e --- /dev/null +++ b/pythia14m_layer5_exp4/history.json @@ -0,0 +1,102 @@ +{ + "loss": [ + 0.002171335807189509, + 0.0004760217304321582, + 0.00029963853915857203 + ], + "recon_loss": [ + 0.0013176010684409186, + 0.0002397540533449184, + 0.00011932067944890448 + ], + "l1_loss": [ + 0.04776091468748645, + 0.06832094815054303, + 0.07560518625684204 + ], + "sparsity": [ + 127.99990193597202, + 127.99837555859352, + 127.99220852583592 + ], + "recon_contribution": [ + 0.0013176010684409186, + 0.0002397540533449184, + 0.00011932067944890448 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 1.8027577064040848e-05 + ], + "logit_kl": [ + 0.08537347564734045, + 0.023626768219321897, + 0.016229028596028098 + ], + "kl_contribution": [ + 0.0008537347380190143, + 0.00023626767695023788, + 0.0001622902823324205 + ], + "dead_features": [ + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.001085616160162565, + 0.00044631636424370105, + 0.00017706250308823795 + ], + "val_recon_loss": [ + 0.0005686862940138595, + 0.0002042529623299786, + 9.947409705241557e-05 + ], + "val_l1_loss": [ + 0.05804161652279463, + 0.07236223318857368, + 0.08036526477049828 + ], + "val_sparsity": [ + 127.99999902596099, + 127.99999862386659, + 127.9999893548975 + ], + "val_recon_contribution": [ + 0.0005686862940138595, + 0.0002042529623299786, + 9.947409705241557e-05 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.05169298785608744, + 0.02420634076909849, + 0.007758840788109249 + ], + "val_kl_contribution": [ + 0.0005169298669883664, + 0.00024206340252859213, + 7.758840610276382e-05 + ] +} \ No newline at end of file diff --git a/pythia14m_layer5_exp4/weights.pt b/pythia14m_layer5_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..60172812919b75bbd8a4a5fd1969a9dba50a4b72 --- /dev/null +++ b/pythia14m_layer5_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:965f5bbd4de6b75611e9efa059e2b89dd8daf2885f12722f64f739f4d40786bc +size 532117 diff --git a/pythia160m_layer0_exp4/config.json b/pythia160m_layer0_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..5e649ac69afb5f1a7f9f55f080dcfc2c818614ee --- /dev/null +++ b/pythia160m_layer0_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 768, "d_sae": 3072, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "0", "hook_name": "hook_resid_post", "hook_spec": "blocks.0.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia160m_layer0_exp4/history.json b/pythia160m_layer0_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..eb2f4a0a49b328e87ce3f78bc9bebdc8ca5732e4 --- /dev/null +++ b/pythia160m_layer0_exp4/history.json @@ -0,0 +1,102 @@ +{ + "loss": [ + 0.01408560986370862, + 0.007514906397444151, + 0.008009068454698134 + ], + "recon_loss": [ + 0.011892832059818668, + 0.006303946278849933, + 0.006514605752281135 + ], + "l1_loss": [ + 0.016570691592661904, + 0.015300610695718503, + 0.014500189677108916 + ], + "sparsity": [ + 125.82674106935255, + 119.6446358557461, + 113.00992185631577 + ], + "recon_contribution": [ + 0.011892832059818668, + 0.006303946278849933, + 0.006514605752281135 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0 + ], + "logit_kl": [ + 0.21927778604788845, + 0.12109601487766723, + 0.1494462743350843 + ], + "kl_contribution": [ + 0.002192777817234156, + 0.0012109601214254071, + 0.0014944627094256762 + ], + "dead_features": [ + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.01905087376617774, + 0.01948434958550584, + 0.017442850234150764 + ], + "val_recon_loss": [ + 0.016354236604827415, + 0.016351540761808697, + 0.014324624026314837 + ], + "val_l1_loss": [ + 0.017131587560272738, + 0.016653423454681057, + 0.017019785634650206 + ], + "val_sparsity": [ + 127.36297070191844, + 126.85029163633631, + 126.02939505339806 + ], + "val_recon_contribution": [ + 0.016354236604827415, + 0.016351540761808697, + 0.014324624026314837 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.2696637213991951, + 0.313280890024716, + 0.3118226270520858 + ], + "val_kl_contribution": [ + 0.0026966371536131088, + 0.003132808827268164, + 0.003118226198313203 + ] +} \ No newline at end of file diff --git a/pythia160m_layer0_exp4/weights.pt b/pythia160m_layer0_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..857437ecd0f7ec88f36758076483fa538109354e --- /dev/null +++ b/pythia160m_layer0_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:219eb0dcd5a5ff3fcd2515f3d36465b84faca3614459ee4315571b5a3e325507 +size 18907797 diff --git a/pythia160m_layer11_exp4/config.json b/pythia160m_layer11_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8ee920dba27b51ed8aeb13244b608f24406ef8a3 --- /dev/null +++ b/pythia160m_layer11_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 768, "d_sae": 3072, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "11", "hook_name": "hook_resid_post", "hook_spec": "blocks.11.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia160m_layer11_exp4/history.json b/pythia160m_layer11_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..8d709141f34a05ac727b665f73d320edf5aaf426 --- /dev/null +++ b/pythia160m_layer11_exp4/history.json @@ -0,0 +1,102 @@ +{ + "loss": [ + 0.31261095851659776, + 0.2187371599532309, + 0.19628824485199792 + ], + "recon_loss": [ + 0.29276702486130657, + 0.20466143189847064, + 0.1841151620812562 + ], + "l1_loss": [ + 0.15303128357647228, + 0.15231893757472234, + 0.13958976851130017 + ], + "sparsity": [ + 127.85488501308727, + 127.70968831899215, + 127.59238116724961 + ], + "recon_contribution": [ + 0.29276702486130657, + 0.20466143189847064, + 0.1841151620812562 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.00677939847232077, + 0.006443772708433268, + 0.00583030989418934 + ], + "logit_kl": [ + 1.3064536393094226, + 0.7631955589161438, + 0.6342773145964357 + ], + "kl_contribution": [ + 0.01306453608910927, + 0.00763195541443392, + 0.006342773007222635 + ], + "dead_features": [ + 8, + 3, + 2 + ], + "dead_feature_percentage": [ + 0.2604166666666667, + 0.09765625, + 0.06510416666666667 + ], + "val_loss": [ + 0.3203053327649324, + 0.3009165501972405, + 0.2737972388817245 + ], + "val_recon_loss": [ + 0.29953614784875693, + 0.2827136098942386, + 0.25727858912420665 + ], + "val_l1_loss": [ + 0.1699585138065277, + 0.15156140507547403, + 0.15585914083541294 + ], + "val_sparsity": [ + 127.99982987542337, + 127.99906710759262, + 127.99931296444134 + ], + "val_recon_contribution": [ + 0.29953614784875693, + 0.2827136098942386, + 0.25727858912420665 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.009392711516766825, + 0.008806341445332504, + 0.00810368043997775 + ], + "val_logit_kl": [ + 1.1376474124278515, + 0.9396599215273151, + 0.8414969695370134 + ], + "val_kl_contribution": [ + 0.011376473860784695, + 0.009396599003208399, + 0.00841496950688365 + ] +} \ No newline at end of file diff --git a/pythia160m_layer11_exp4/weights.pt b/pythia160m_layer11_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..a3f8a1da6cd5999c46e69c88ec8e9a6feb9628f3 --- /dev/null +++ b/pythia160m_layer11_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4020595dfc8ac3447accfe37dc3e463644822c9ba780221f1ab0016b06462ee7 +size 18907797 diff --git a/pythia160m_layer2_exp4/config.json b/pythia160m_layer2_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b4ccfdf3f46891a0d8e729aee7f3e4d99d7c2788 --- /dev/null +++ b/pythia160m_layer2_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 768, "d_sae": 3072, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "2", "hook_name": "hook_resid_post", "hook_spec": "blocks.2.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia160m_layer2_exp4/history.json b/pythia160m_layer2_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..d3510a5f927d1344baffb48ab042121106895fc3 --- /dev/null +++ b/pythia160m_layer2_exp4/history.json @@ -0,0 +1,102 @@ +{ + "loss": [ + 0.022805505420151008, + 0.013441352083404757, + 0.011892197134454741 + ], + "recon_loss": [ + 0.02104288216927374, + 0.012321489272309707, + 0.010873594475897396 + ], + "l1_loss": [ + 0.02589880668108358, + 0.0255978409492341, + 0.02548265174335363 + ], + "sparsity": [ + 127.99894944275317, + 127.82450345616762, + 127.5587693298755 + ], + "recon_contribution": [ + 0.02104288216927374, + 0.012321489272309707, + 0.010873594475897396 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0003576286187887249, + 0.0003886501407559619, + 0.00034346553040858434 + ], + "logit_kl": [ + 0.1404994673342729, + 0.07312126973403149, + 0.06751371551761214 + ], + "kl_contribution": [ + 0.001404994644234939, + 0.0007312126809114996, + 0.0006751371405426363 + ], + "dead_features": [ + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.03972662800457011, + 0.03050375607654442, + 0.02940352670289703 + ], + "val_recon_loss": [ + 0.03624709893205923, + 0.027982315392028376, + 0.026998186516356493 + ], + "val_l1_loss": [ + 0.0314848783877364, + 0.03010873408796394, + 0.02960533726180827 + ], + "val_sparsity": [ + 127.999844010866, + 127.9903848930208, + 127.97814808268895 + ], + "val_recon_contribution": [ + 0.03624709893205923, + 0.027982315392028376, + 0.026998186516356493 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0011356308797065514, + 0.0008778145541581398, + 0.0008475123470171082 + ], + "val_logit_kl": [ + 0.2343898291541511, + 0.16436261869332136, + 0.15578279077712176 + ], + "val_kl_contribution": [ + 0.0023438982390887506, + 0.0016436261488677046, + 0.0015578278736465298 + ] +} \ No newline at end of file diff --git a/pythia160m_layer2_exp4/weights.pt b/pythia160m_layer2_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..eed20924ef274ae38deefa1ee5b094501f2a7f49 --- /dev/null +++ b/pythia160m_layer2_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df96f23810e4560648c49d7349f8860cb0ecfe9af5208dbe0b4dfc33fe248aac +size 18907797 diff --git a/pythia160m_layer5_exp4/config.json b/pythia160m_layer5_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..5a32d879b820b5051009f33d6a094e0c0847f6fc --- /dev/null +++ b/pythia160m_layer5_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 768, "d_sae": 3072, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "5", "hook_name": "hook_resid_post", "hook_spec": "blocks.5.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia160m_layer5_exp4/history.json b/pythia160m_layer5_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..408721037c33057419062988c7df46e8aaeeb1f0 --- /dev/null +++ b/pythia160m_layer5_exp4/history.json @@ -0,0 +1,202 @@ +{ + "loss": [ + 0.040373627999227266, + 0.0235105927501406, + 0.021295735231112865, + 0.02013097347653642, + 0.019527294850420384, + 0.02013132228344983, + 0.020482758718992577, + 0.01904547135662414 + ], + "recon_loss": [ + 0.03780413387710748, + 0.022013843651911637, + 0.01985880850947329, + 0.018754747890088023, + 0.018214199935295144, + 0.01880487274365867, + 0.01909509737513402, + 0.0177838154010424 + ], + "l1_loss": [ + 0.043025053436962925, + 0.03850345717248868, + 0.03658427574044588, + 0.035529070059914576, + 0.03519518073318767, + 0.034862890761016174, + 0.03476217700307872, + 0.035215672171440254 + ], + "sparsity": [ + 127.99999003507654, + 127.99983514279735, + 127.99785971090097, + 127.990553896119, + 127.9786807936065, + 127.85034133495928, + 127.71922843569801, + 127.97289638389536 + ], + "recon_contribution": [ + 0.03780413387710748, + 0.022013843651911637, + 0.01985880850947329, + 0.018754747890088023, + 0.018214199935295144, + 0.01880487274365867, + 0.01909509737513402, + 0.0177838154010424 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0007852440172502277, + 0.0006900960318173985, + 0.000623730818291993, + 0.0005906092030309825, + 0.000575056742301819, + 0.0005994714941864288, + 0.0006408378134598816, + 0.0005992114398593628 + ], + "logit_kl": [ + 0.17842501832463709, + 0.0806653115733647, + 0.0813195915373207, + 0.07856163996316137, + 0.07380382034065974, + 0.07269781013811324, + 0.0746823544012142, + 0.0662444544582292 + ], + "kl_contribution": [ + 0.0017842501426451416, + 0.0008066530980793031, + 0.000813195896873466, + 0.0007856163825462794, + 0.0007380381860884324, + 0.0007269780845880243, + 0.0007468235271517187, + 0.0006624445293041669 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.050047308106252186, + 0.045878467521418824, + 0.04193598931069958, + 0.03932654552146457, + 0.03896158611771397, + 0.03918718963470399, + 0.03783226667902729, + 0.03736098678865112 + ], + "val_recon_loss": [ + 0.04597878260074543, + 0.04146085552434452, + 0.038738831965317515, + 0.036382041950268895, + 0.036052692481573737, + 0.03624962614538134, + 0.03494560202087219, + 0.0344847117835718 + ], + "val_l1_loss": [ + 0.06750534617563617, + 0.05467167289466702, + 0.04928356658187381, + 0.045380183623185674, + 0.04458471999872519, + 0.042920941640587996, + 0.04257036754182186, + 0.0421116140295879 + ], + "val_sparsity": [ + 127.99998989508278, + 127.99992429893197, + 127.99980824060316, + 127.99982074303293, + 127.9998119330845, + 127.99730209111355, + 127.99881933689443, + 127.99898187888262 + ], + "val_recon_contribution": [ + 0.04597878260074543, + 0.04146085552434452, + 0.038738831965317515, + 0.036382041950268895, + 0.036052692481573737, + 0.03624962614538134, + 0.03494560202087219, + 0.0344847117835718 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0014391526589458954, + 0.001297814393837904, + 0.0012136513375072158, + 0.0011413065298735313, + 0.001132682456109443, + 0.001155864660366741, + 0.0011376148454526574, + 0.001125563739527724 + ], + "val_logit_kl": [ + 0.26293730226376466, + 0.31197976425604995, + 0.19835060795524043, + 0.18031970656336613, + 0.17762112186303655, + 0.1781698907988975, + 0.17490498527589915, + 0.1750711383654824 + ], + "val_kl_contribution": [ + 0.0026293729636507096, + 0.0031197975703433217, + 0.0019835060334671734, + 0.001803197027077732, + 0.0017762111800506295, + 0.001781698869566364, + 0.0017490498144681114, + 0.0017507113439752134 + ] +} \ No newline at end of file diff --git a/pythia160m_layer5_exp4/weights.pt b/pythia160m_layer5_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..8744cf328ac27207f5346482a20b70ddcd20e51e --- /dev/null +++ b/pythia160m_layer5_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89b5e158449e476f097c7c038296ee4966b8369926fa44f2ca7d92e389e499ff +size 18907797 diff --git a/pythia160m_layer8_exp4/config.json b/pythia160m_layer8_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d38942415045ad82aa2c2c12dbef1cccda87cbbf --- /dev/null +++ b/pythia160m_layer8_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 768, "d_sae": 3072, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "8", "hook_name": "hook_resid_post", "hook_spec": "blocks.8.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia160m_layer8_exp4/history.json b/pythia160m_layer8_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..cd3935b7ac707c45947dde9c64f3e450ac4b8795 --- /dev/null +++ b/pythia160m_layer8_exp4/history.json @@ -0,0 +1,62 @@ +{ + "loss": [ + 0.06083375318165945 + ], + "recon_loss": [ + 0.05789181708934761 + ], + "l1_loss": [ + 0.05494845615965979 + ], + "sparsity": [ + 127.99917661705796 + ], + "recon_contribution": [ + 0.05789181708934761 + ], + "l1_contribution": [ + 0.0 + ], + "aux_loss": [ + 0.0012162892464279088 + ], + "logit_kl": [ + 0.17256469051702086 + ], + "kl_contribution": [ + 0.0017256468694441679 + ], + "dead_features": [ + 0 + ], + "dead_feature_percentage": [ + 0.0 + ], + "val_loss": [ + 0.0773609928601379 + ], + "val_recon_loss": [ + 0.07241062774756528 + ], + "val_l1_loss": [ + 0.07569787406248955 + ], + "val_sparsity": [ + 127.99990451035242 + ], + "val_recon_contribution": [ + 0.07241062774756528 + ], + "val_l1_contribution": [ + 0.0 + ], + "val_aux_loss": [ + 0.0022719658701886927 + ], + "val_logit_kl": [ + 0.26783993265298894 + ], + "val_kl_contribution": [ + 0.0026783992648019943 + ] +} \ No newline at end of file diff --git a/pythia160m_layer8_exp4/weights.pt b/pythia160m_layer8_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..b79517a6fb79718dc5a79f2e5a4b373fa1cb2515 --- /dev/null +++ b/pythia160m_layer8_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad0bbf1910051156ccbf579f7681eb36ca30f1ed828526c0aea5c17b141e0e47 +size 18907797 diff --git a/pythia31m_layer0_exp4/config.json b/pythia31m_layer0_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..3c40ef0711c49c837a79be59ea7589e2bfee1fbd --- /dev/null +++ b/pythia31m_layer0_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 256, "d_sae": 1024, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "0", "hook_name": "hook_resid_post", "hook_spec": "blocks.0.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia31m_layer0_exp4/history.json b/pythia31m_layer0_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..a30920b1e501f00d851da32e471f56939ee73cb7 --- /dev/null +++ b/pythia31m_layer0_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.004300681019845012, + 0.0020750122649355125, + 0.0016903404607340362, + 0.0014801529093946375, + 0.0013317227194139805, + 0.0011789969790506323, + 0.0010967958886928504, + 0.0010386491676614182, + 0.0009632098909802126, + 0.0009150082608501084 + ], + "recon_loss": [ + 0.002484193599711894, + 0.0012011405934867856, + 0.0009437641815822014, + 0.0007954730217616765, + 0.0006917270887436044, + 0.0005960185320305464, + 0.0005401509802396783, + 0.0004984021615251788, + 0.00045303893673293694, + 0.0004234032033370705 + ], + "l1_loss": [ + 0.018314964766740906, + 0.018118779445422, + 0.018788140240089863, + 0.019709822628816112, + 0.02073845377499525, + 0.021730719046319028, + 0.02260436980009836, + 0.0232727119070351, + 0.024081100204354406, + 0.024300629562088127 + ], + "sparsity": [ + 127.99999921075229, + 127.99115291394685, + 127.54464618523627, + 126.45392124016792, + 125.14519540194807, + 124.08716318092415, + 123.38990746305988, + 122.22088900668219, + 121.69541614068615, + 120.11455145418752 + ], + "recon_contribution": [ + 0.002484193599711894, + 0.0012011405934867856, + 0.0009437641815822014, + 0.0007954730217616765, + 0.0006917270887436044, + 0.0005960185320305464, + 0.0005401509802396783, + 0.0004984021615251788, + 0.00045303893673293694, + 0.0004234032033370705 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "logit_kl": [ + 0.18164874549732019, + 0.08738716913491974, + 0.07465762978321627, + 0.06846799008115557, + 0.06399956472326213, + 0.05829784584429433, + 0.05566449223691344, + 0.05402470177238952, + 0.0510170965585621, + 0.04916050680110167 + ], + "kl_contribution": [ + 0.0018164874141644552, + 0.000873871671026167, + 0.0007465762809477155, + 0.0006846798862332306, + 0.0006399956324398467, + 0.0005829784454883055, + 0.0005566449099585428, + 0.0005402470059513694, + 0.0005101709539303556, + 0.0004916050573413729 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.00479095759829058, + 0.003840290768842613, + 0.003157867785275903, + 0.003035797289218229, + 0.0024278055505148667, + 0.0021996161770193984, + 0.0020500354380638664, + 0.0019430801729488946, + 0.0017236777858620003, + 0.0016295625032014555 + ], + "val_recon_loss": [ + 0.0024979233128068585, + 0.001962823910870463, + 0.0016170786069321203, + 0.001491002234062273, + 0.001214182953471631, + 0.0010834264972208695, + 0.0009939830791358274, + 0.0009279896961269202, + 0.0008203235080974899, + 0.0007575288346015276 + ], + "val_l1_loss": [ + 0.018823917124498227, + 0.019756025549271344, + 0.021182555520905352, + 0.0226428556838813, + 0.024241922446928382, + 0.025526566652548953, + 0.02647227614603727, + 0.027599845347176563, + 0.02824329997461076, + 0.028785429658075554 + ], + "val_sparsity": [ + 127.99999901600906, + 127.99827320677491, + 127.94499762289963, + 127.8050213778268, + 127.68573759600426, + 127.56540580146644, + 127.43755226932149, + 127.32129372606269, + 127.1562900456594, + 126.93422992785989 + ], + "val_recon_contribution": [ + 0.0024979233128068585, + 0.001962823910870463, + 0.0016170786069321203, + 0.001491002234062273, + 0.001214182953471631, + 0.0010834264972208695, + 0.0009939830791358274, + 0.0009279896961269202, + 0.0008203235080974899, + 0.0007575288346015276 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.22930343301592904, + 0.18774669021909396, + 0.1540789210679962, + 0.15447950859187842, + 0.12136226248581446, + 0.11161897070604385, + 0.10560523832621194, + 0.10150904998142647, + 0.09033542977042029, + 0.0872033689675225 + ], + "val_kl_contribution": [ + 0.002293034278822354, + 0.0018774668614614373, + 0.0015407891760704587, + 0.0015447950534113123, + 0.0012136225976247838, + 0.0011161896820982866, + 0.0010560523594302853, + 0.0010150904775621263, + 0.0009033542776323405, + 0.0008720336701595339 + ] +} \ No newline at end of file diff --git a/pythia31m_layer0_exp4/weights.pt b/pythia31m_layer0_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..c2ba021a429ef4106c8824ebf37c286e3a7e8239 --- /dev/null +++ b/pythia31m_layer0_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6d4520603b2eed5092dde9c9895472ba4cc24f8e82c87b9f7fda1f18301def0 +size 2110101 diff --git a/pythia31m_layer2_exp4/config.json b/pythia31m_layer2_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..0b9c79aa954e87a3a2c565f69d5f35764298742a --- /dev/null +++ b/pythia31m_layer2_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 256, "d_sae": 1024, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "2", "hook_name": "hook_resid_post", "hook_spec": "blocks.2.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia31m_layer2_exp4/history.json b/pythia31m_layer2_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee62363703b006002c20359f9f62e3bd1b1c068 --- /dev/null +++ b/pythia31m_layer2_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.004840253223749571, + 0.002278674162434965, + 0.0018178231114865586, + 0.0015814056679794054, + 0.0014277329646112623, + 0.0013412063425413574, + 0.0012756717862889252, + 0.0012246038152108614, + 0.0011730338734665044, + 0.001148436724198999 + ], + "recon_loss": [ + 0.0034238550782004227, + 0.00162121779585389, + 0.0012685014878863641, + 0.0010882465497057332, + 0.0009699914813170146, + 0.0009057100239727889, + 0.0008553544285412651, + 0.0008149448384840579, + 0.0007787483570070011, + 0.000759457324907444 + ], + "l1_loss": [ + 0.023966038391710735, + 0.028137029547535032, + 0.031138668188355364, + 0.03273184595632953, + 0.033466166094919085, + 0.034175343001886745, + 0.03464115422536803, + 0.03475952192890828, + 0.034991624315494095, + 0.03512168232818048 + ], + "sparsity": [ + 127.99999894766972, + 127.99999879535876, + 127.99998512891156, + 127.99940252563698, + 127.98986901561923, + 127.97287960017874, + 127.95130734867712, + 127.93205198921439, + 127.92013491649593, + 127.90370445459162 + ], + "recon_contribution": [ + 0.0034238550782004227, + 0.00162121779585389, + 0.0012685014878863641, + 0.0010882465497057332, + 0.0009699914813170146, + 0.0009057100239727889, + 0.0008553544285412651, + 0.0008149448384840579, + 0.0007787483570070011, + 0.000759457324907444 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "logit_kl": [ + 0.14163981762439176, + 0.06574563817636613, + 0.05493216343670272, + 0.04931591293975684, + 0.045774149370401716, + 0.043549632888537464, + 0.04203173656442529, + 0.04096589860474241, + 0.03942855264139641, + 0.03889794090252282 + ], + "kl_contribution": [ + 0.0014163981409009863, + 0.0006574563673997852, + 0.000549321622081619, + 0.0004931591184849525, + 0.000457741483624373, + 0.00043549631904394866, + 0.00042031735608382944, + 0.0004096589767796236, + 0.0003942855173970586, + 0.00038897940029513544 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.00502709000156541, + 0.0039007034772147614, + 0.0032288803249304874, + 0.002927257581174096, + 0.002804508758307038, + 0.00264317928367192, + 0.002608600578615129, + 0.002434029965111639, + 0.0023914756387434935, + 0.002282403783945047 + ], + "val_recon_loss": [ + 0.003457292411298737, + 0.002653297891127255, + 0.0022225459569390122, + 0.00200313399450686, + 0.001894875585002576, + 0.0017899646719077688, + 0.0017393108522460257, + 0.001639146436600913, + 0.001598493763660973, + 0.0015431521925652275 + ], + "val_l1_loss": [ + 0.029255775750008203, + 0.03409305798544546, + 0.03660890540881226, + 0.03787013485547253, + 0.03855880506532935, + 0.03901114673451009, + 0.039228656138889796, + 0.03985278617817364, + 0.03992428533313331, + 0.04013821589409439 + ], + "val_sparsity": [ + 127.99999901600906, + 127.99999901600906, + 127.99998669533379, + 127.99988733303752, + 127.99878762615042, + 127.99467266612872, + 127.99287220524398, + 127.99659067928305, + 127.99316111052718, + 127.99029159199422 + ], + "val_recon_contribution": [ + 0.003457292411298737, + 0.002653297891127255, + 0.0022225459569390122, + 0.00200313399450686, + 0.001894875585002576, + 0.0017899646719077688, + 0.0017393108522460257, + 0.001639146436600913, + 0.001598493763660973, + 0.0015431521925652275 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.15697976241360242, + 0.1247405610626714, + 0.10063343908938899, + 0.09241236033735223, + 0.09096331913683979, + 0.0853214634381079, + 0.08692897454269576, + 0.07948835468855259, + 0.07929818935820886, + 0.07392516069547303 + ], + "val_kl_contribution": [ + 0.001569797589473653, + 0.0012474055831268984, + 0.001006334367727135, + 0.0009241235837066287, + 0.0009096331720620641, + 0.000853214615332741, + 0.0008692897259197252, + 0.000794883529224444, + 0.0007929818747124443, + 0.0007392515909833094 + ] +} \ No newline at end of file diff --git a/pythia31m_layer2_exp4/weights.pt b/pythia31m_layer2_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..b6371306a5ab62fcaf31efabd64a8a40c0b1e11f --- /dev/null +++ b/pythia31m_layer2_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f34741f7fc57ee1eaeb9f406d1bcbd75be26738362375cf94b85f4f244bc225b +size 2110101 diff --git a/pythia31m_layer5_exp4/config.json b/pythia31m_layer5_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..1f6dae89977399664f2d488d51c7f2f1935ed9ef --- /dev/null +++ b/pythia31m_layer5_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 256, "d_sae": 1024, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "5", "hook_name": "hook_resid_post", "hook_spec": "blocks.5.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia31m_layer5_exp4/history.json b/pythia31m_layer5_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..4cfe847606c473e7de30b44e5b2c1bcb9be7173f --- /dev/null +++ b/pythia31m_layer5_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.01302572470217497, + 0.006334375048786429, + 0.005149282265527211, + 0.004496611818289232, + 0.004148950476407288, + 0.003787454145203282, + 0.003564906887585017, + 0.0033458373704781507, + 0.003125993215361485, + 0.0030131820160827975 + ], + "recon_loss": [ + 0.01142719742367061, + 0.005590697214642908, + 0.004537759655951345, + 0.003947360160720016, + 0.003594230861316846, + 0.0032965857957104414, + 0.0030951943612334653, + 0.0028815273986326465, + 0.0027076934539068983, + 0.002581914819645173 + ], + "l1_loss": [ + 0.05338030695212082, + 0.06104766827975881, + 0.0636382561596195, + 0.06481172291801976, + 0.0648203661830964, + 0.06476055569067841, + 0.0642494309673292, + 0.0634408291638871, + 0.06275553809996093, + 0.06169423727342308 + ], + "sparsity": [ + 127.99999894766972, + 127.9999989684394, + 127.99999569375372, + 127.99996844393804, + 127.9994392810413, + 127.99841874221275, + 127.99723561835592, + 127.99646395563863, + 127.9961876427846, + 127.9972310005645 + ], + "recon_contribution": [ + 0.01142719742367061, + 0.005590697214642908, + 0.004537759655951345, + 0.003947360160720016, + 0.003594230861316846, + 0.0032965857957104414, + 0.0030951943612334653, + 0.0028815273986326465, + 0.0027076934539068983, + 0.002581914819645173 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 1.7302230243679277e-05, + 3.6648317296062165e-05, + 2.1797268164266805e-05, + 3.2166564811902974e-05, + 1.429139824549054e-05, + 2.4387967948830974e-05 + ], + "logit_kl": [ + 0.15985273175320153, + 0.07436778483415418, + 0.061152261987788274, + 0.05492516664007236, + 0.05374173960984835, + 0.04542200427887769, + 0.04479152615775376, + 0.0432143418233294, + 0.04040083714918032, + 0.040687923729521125 + ], + "kl_contribution": [ + 0.0015985272828884212, + 0.0007436778312912397, + 0.0006115226066179452, + 0.0005492516540830951, + 0.0005374173841865128, + 0.00045422003226280346, + 0.00044791525210638, + 0.0004321434086908294, + 0.0004040083621791057, + 0.0004068792279936059 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.012593392951398764, + 0.010081298126552346, + 0.00909839349830156, + 0.009864159005370117, + 0.007821473611060269, + 0.007082399969910365, + 0.006772226295971078, + 0.007223451139361383, + 0.005915704346959776, + 0.005833740660381063 + ], + "val_recon_loss": [ + 0.010802852455103646, + 0.008644133223411876, + 0.007623522123068124, + 0.0071035392608216116, + 0.006411265383485526, + 0.005986284447891416, + 0.0056099123008495075, + 0.0055146071770999045, + 0.0049199635288986856, + 0.004769040122975026 + ], + "val_l1_loss": [ + 0.05918895065459633, + 0.06397366729462613, + 0.06503240367927192, + 0.06557243518348611, + 0.06482489162093395, + 0.06454214074739427, + 0.06382626404378439, + 0.061668518981426876, + 0.0627917640188626, + 0.06102580897557421 + ], + "val_sparsity": [ + 127.99999901600906, + 127.99999901600906, + 127.99999901600906, + 127.99997027238643, + 127.99944398968356, + 127.99947956580665, + 127.99942305562281, + 127.9930292557414, + 127.99932314589498, + 127.99888194376074 + ], + "val_recon_contribution": [ + 0.010802852455103646, + 0.008644133223411876, + 0.007623522123068124, + 0.0071035392608216116, + 0.006411265383485526, + 0.005986284447891416, + 0.0056099123008495075, + 0.0055146071770999045, + 0.0049199635288986856, + 0.004769040122975026 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0002775527701288509, + 0.0, + 0.0, + 0.00019326450697037026, + 0.00017862027186734794, + 0.00016590934168533858 + ], + "val_logit_kl": [ + 0.17905405332265822, + 0.1437164928144915, + 0.14748713988929202, + 0.27606198113144365, + 0.1132655481382406, + 0.10961155515911902, + 0.11623140145346969, + 0.1515579490459691, + 0.08171205628971205, + 0.08987912293235949 + ], + "val_kl_contribution": [ + 0.001790540493281641, + 0.0014371648950516657, + 0.0014748713662458767, + 0.0027606197468746964, + 0.0011326554564942692, + 0.0010961155265391627, + 0.0011623139878522205, + 0.0015155794566855017, + 0.000817120544429273, + 0.0008987912084618846 + ] +} \ No newline at end of file diff --git a/pythia31m_layer5_exp4/weights.pt b/pythia31m_layer5_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..0b1a477144f27696a38c5c16475267fecc6deeb3 --- /dev/null +++ b/pythia31m_layer5_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b33bf87955234d9181b1acd282cac084c9ebb5a1615370cdd87107ea3f1cf09 +size 2110101 diff --git a/pythia410m_layer0_exp4/config.json b/pythia410m_layer0_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..43c694f0bc3a348ace39e6494ba26194ddc01139 --- /dev/null +++ b/pythia410m_layer0_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "0", "hook_name": "hook_resid_post", "hook_spec": "blocks.0.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer0_exp4/history.json b/pythia410m_layer0_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..bd7e9e5ac604b35b5547eefea16aaad5540a366b --- /dev/null +++ b/pythia410m_layer0_exp4/history.json @@ -0,0 +1,162 @@ +{ + "loss": [ + 0.02322444988299709, + 0.018646012573120964, + 0.01737235312412633, + 0.015005767014458228, + 0.015878045804901708, + 0.01305246714723572 + ], + "recon_loss": [ + 0.02027805474271555, + 0.016173136122854335, + 0.014859554402406806, + 0.012985716067138761, + 0.013644954730323879, + 0.011271215657941902 + ], + "l1_loss": [ + 0.015994312166035203, + 0.013097935160976952, + 0.012171232466925672, + 0.01153450303169275, + 0.010803141293983007, + 0.010980826322734694 + ], + "sparsity": [ + 122.42357081640004, + 101.26598719027427, + 81.48973336193825, + 67.09398575486809, + 54.792241452177294, + 51.434536035610414 + ], + "recon_contribution": [ + 0.02027805474271555, + 0.016173136122854335, + 0.014859554402406806, + 0.012985716067138761, + 0.013644954730323879, + 0.011271215657941902 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.000324454694522478, + 0.0005382559986946366, + 0.0005608199303003134, + 0.00048689568389824906, + 0.0005315753970913291, + 0.000449218279342203 + ], + "logit_kl": [ + 0.2621940491085504, + 0.19346204612525805, + 0.19519788424354284, + 0.1533155285579215, + 0.1701515712614717, + 0.1332033246496993 + ], + "kl_contribution": [ + 0.002621940434891342, + 0.0019346204188037722, + 0.0019519787990913207, + 0.0015331552511537675, + 0.0017015156734655754, + 0.001332033215689189 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.03434971954482271, + 0.042083952062657914, + 0.03296107299133343, + 0.032812064094659366, + 0.032438541241227116, + 0.028229163100656082 + ], + "val_recon_loss": [ + 0.029236371849817155, + 0.034351466901038435, + 0.027217990994031986, + 0.027579497245530927, + 0.02698679796455158, + 0.023628788665512483 + ], + "val_l1_loss": [ + 0.01531838950595697, + 0.014435824888188398, + 0.013629610239799349, + 0.013034051475451513, + 0.012736345495756161, + 0.013262153712305656 + ], + "val_sparsity": [ + 126.35284980867517, + 113.19826088142308, + 94.04735224098576, + 81.22015007882159, + 69.82908706110426, + 67.10531562656124 + ], + "val_recon_contribution": [ + 0.029236371849817155, + 0.034351466901038435, + 0.027217990994031986, + 0.027579497245530927, + 0.02698679796455158, + 0.023628788665512483 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0009785492232676361, + 0.0011667576835746548, + 0.000938603805312176, + 0.0009437479296393874, + 0.0009271282813519617, + 0.0008241019077178663 + ], + "val_logit_kl": [ + 0.4134798564934178, + 0.6565727608169225, + 0.4804478284277899, + 0.4288818996834625, + 0.4524615113333592, + 0.37762726124871976 + ], + "val_kl_contribution": [ + 0.004134798473681703, + 0.0065657274607888024, + 0.0048044781823761455, + 0.004288818901995013, + 0.004524615013399421, + 0.003776272530757929 + ] +} \ No newline at end of file diff --git a/pythia410m_layer0_exp4/weights.pt b/pythia410m_layer0_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..dc34e939e2dfb31e5563f2b7dcafc40955d23ac0 --- /dev/null +++ b/pythia410m_layer0_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0048d4bdfd4a04d9127c7e85a5f26de7c50b118df12388d3106722fcafe0be1 +size 33598101 diff --git a/pythia410m_layer12_exp4/config.json b/pythia410m_layer12_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b5233045dbb0f9cc6c699980d88a82488083aca6 --- /dev/null +++ b/pythia410m_layer12_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "12", "hook_name": "hook_resid_post", "hook_spec": "blocks.12.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer12_exp4/history.json b/pythia410m_layer12_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..e941fa3b51f44e1d08cbcccfca0d6dbee5467b68 --- /dev/null +++ b/pythia410m_layer12_exp4/history.json @@ -0,0 +1,222 @@ +{ + "loss": [ + 0.0633780482815014, + 0.040449679832560724, + 0.03699899633362646, + 0.03426451659879686, + 0.03293758990266741, + 0.03207839903093793, + 0.03173348016212078, + 0.031233223596293675, + 0.030805410166522877 + ], + "recon_loss": [ + 0.06043631105367669, + 0.038418971138588824, + 0.035141831789564634, + 0.03255485847445303, + 0.031354127378251995, + 0.030511992129525965, + 0.03020807131643182, + 0.02974190070849933, + 0.02936934014974997 + ], + "l1_loss": [ + 0.04254397630569831, + 0.03591604542389458, + 0.03335641302631495, + 0.032121884312054746, + 0.031592240415094335, + 0.03115089382712616, + 0.0310382004413152, + 0.03091504763450115, + 0.030861117897349022 + ], + "sparsity": [ + 127.99999706455237, + 127.9996121193666, + 127.99637660876377, + 127.98658852135854, + 127.96970221697744, + 127.951625713203, + 127.9493801200022, + 127.94733452147884, + 127.94911084114531 + ], + "recon_contribution": [ + 0.06043631105367669, + 0.038418971138588824, + 0.035141831789564634, + 0.03255485847445303, + 0.031354127378251995, + 0.030511992129525965, + 0.03020807131643182, + 0.02974190070849933, + 0.02936934014974997 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0013407035204427683, + 0.001206279968894205, + 0.0011090461909263893, + 0.001031540746801869, + 0.0009955380241685131, + 0.0009693856348674246, + 0.0009607186671241843, + 0.0009464948740661037, + 0.0009347905664793055 + ], + "logit_kl": [ + 0.16010338188860626, + 0.08244287678397205, + 0.07481183652206579, + 0.06781174056231976, + 0.05879245127995791, + 0.05970212593953549, + 0.0564690228549936, + 0.05448280248077698, + 0.050127942717639214 + ], + "kl_contribution": [ + 0.001601033781792148, + 0.0008244287497182284, + 0.0007481183488240035, + 0.0006781173903706248, + 0.0005879244992334175, + 0.0005970212459777497, + 0.0005646902159745458, + 0.0005448280124407546, + 0.0005012794155875479 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.13484913517282204, + 0.08673548322172665, + 0.07260407511397066, + 0.09031630709753802, + 0.069086731533803, + 0.06976089404319591, + 0.06714305356618817, + 0.06708424529826598, + 0.06623545442045628 + ], + "val_recon_loss": [ + 0.12502442019018786, + 0.08171022347835101, + 0.06793430633250555, + 0.08208081987802818, + 0.06493967481428035, + 0.06584496814918377, + 0.06334391628806291, + 0.06323459462052392, + 0.0625037070404381 + ], + "val_l1_loss": [ + 0.0801361854501688, + 0.06303330405578457, + 0.05338532578408285, + 0.047394623265672356, + 0.044924426564048874, + 0.042197366175273504, + 0.041195685872431505, + 0.04001485846973939, + 0.03953232180102069 + ], + "val_sparsity": [ + 127.99999902596099, + 127.9999050225306, + 127.99987431430426, + 127.99977505938242, + 127.99917653494128, + 127.99904358034944, + 127.9994777487062, + 127.997718065774, + 127.9978028799609 + ], + "val_recon_contribution": [ + 0.12502442019018786, + 0.08171022347835101, + 0.06793430633250555, + 0.08208081987802818, + 0.06493967481428035, + 0.06584496814918377, + 0.06334391628806291, + 0.06323459462052392, + 0.0625037070404381 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.003909160225321689, + 0.0025626278714785036, + 0.002135797693944836, + 0.0025796899164746306, + 0.002045745794176429, + 0.0020736640022027946, + 0.001995693062206156, + 0.0019915359528347686, + 0.001970262986344236 + ], + "val_logit_kl": [ + 0.5915554641851345, + 0.24626321253831795, + 0.25339712481480303, + 0.5655797486722821, + 0.21013109983773948, + 0.18422620171263834, + 0.18034442895003636, + 0.18581148112418705, + 0.17614844139294697 + ], + "val_kl_contribution": [ + 0.005915554506022539, + 0.0024626320693430332, + 0.0025339711937538835, + 0.005655797356085895, + 0.00210131095507153, + 0.001842261975140703, + 0.0018034442500080125, + 0.00185811476983906, + 0.0017614843742361293 + ] +} \ No newline at end of file diff --git a/pythia410m_layer12_exp4/weights.pt b/pythia410m_layer12_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..2dfd8e73546c77a80d499f3506bbfbb037f5ef1a --- /dev/null +++ b/pythia410m_layer12_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a2255ec8c37ecfcb854ce1ccc53ecab0f64002f18a20d0f865564f16df12a12 +size 33598101 diff --git a/pythia410m_layer16_exp4/config.json b/pythia410m_layer16_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8b54b4a6c8060d6dbba2be55687b77b370999094 --- /dev/null +++ b/pythia410m_layer16_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "16", "hook_name": "hook_resid_post", "hook_spec": "blocks.16.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer16_exp4/history.json b/pythia410m_layer16_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..0003a62682e6c7adf860aedf5ff6b744dc0e77b5 --- /dev/null +++ b/pythia410m_layer16_exp4/history.json @@ -0,0 +1,242 @@ +{ + "loss": [ + 0.08353466329041416, + 0.05369645557594494, + 0.047405824230006895, + 0.044784668386793826, + 0.04229313513422943, + 0.041824697643626625, + 0.04138549756404087, + 0.04064657775593918, + 0.040763516015391764, + 0.040025475279456796 + ], + "recon_loss": [ + 0.07973694055004826, + 0.051114708574796386, + 0.04514499670722794, + 0.042655482153812194, + 0.040345438679987855, + 0.0398980563738047, + 0.03948248920712059, + 0.03881170973519773, + 0.038894988390230006, + 0.038235727643523156 + ], + "l1_loss": [ + 0.04915676139693998, + 0.04172572852729291, + 0.03947106894602603, + 0.03876089097462094, + 0.038228173823875056, + 0.038097299605831095, + 0.038022073062674265, + 0.03790181738403905, + 0.03772651518740693, + 0.03776252770394141 + ], + "sparsity": [ + 127.99964790898096, + 127.99597880368657, + 127.99156529933701, + 127.97838906497574, + 127.96438878331124, + 127.95379145738218, + 127.94487200977582, + 127.93863260854178, + 127.8981907008564, + 127.94249999847689 + ], + "recon_contribution": [ + 0.07973694055004826, + 0.051114708574796386, + 0.04514499670722794, + 0.042655482153812194, + 0.040345438679987855, + 0.0398980563738047, + 0.03948248920712059, + 0.03881170973519773, + 0.038894988390230006, + 0.038235727643523156 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0018474537990520372, + 0.0016171679440222363, + 0.001449510196445488, + 0.0013769177987272793, + 0.0013080102823555938, + 0.0012934218953530932, + 0.0012797989406937332, + 0.0012576378080757257, + 0.0012620125437483273, + 0.001237945258316327 + ], + "logit_kl": [ + 0.1950268866206957, + 0.09645791333287061, + 0.08113173586054419, + 0.07522684408473476, + 0.06396862018021827, + 0.0633219388218938, + 0.06232094896844032, + 0.05772301857148427, + 0.06065151304642107, + 0.05518023526598573 + ], + "kl_contribution": [ + 0.0019502688225744521, + 0.0009645791112404306, + 0.0008113173406807044, + 0.0007522684241525242, + 0.0006396861880489062, + 0.0006332193745150482, + 0.0006232094753783651, + 0.0005772301725555273, + 0.0006065151165585484, + 0.0005518023404650354 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.12729294324079626, + 0.10341824259421112, + 0.09583418694972613, + 0.09145536595301322, + 0.09011536416026854, + 0.08824515689006994, + 0.08703586160595013, + 0.09029735316747105, + 0.08669550285447407, + 0.08668003849658111 + ], + "val_recon_loss": [ + 0.12030711326272528, + 0.09713160396062315, + 0.09056324481300634, + 0.08660473329933704, + 0.08515992711151339, + 0.08358545409687293, + 0.08249965653344751, + 0.08519215440832041, + 0.08217771135284135, + 0.08209042393368436 + ], + "val_l1_loss": [ + 0.08187440372174136, + 0.06497470241739012, + 0.058365051371842176, + 0.053752636876310995, + 0.05160476077265059, + 0.05018334892806493, + 0.04871046773307381, + 0.048060337744182696, + 0.047186675193342555, + 0.04675913578841238 + ], + "val_sparsity": [ + 127.99992907886774, + 127.99915684618121, + 127.99892826591606, + 127.99672018533835, + 127.99829922736313, + 127.99585862530193, + 127.98862459008556, + 127.99453963089509, + 127.99135453372801, + 127.99394974108449 + ], + "val_recon_contribution": [ + 0.12030711326272528, + 0.09713160396062315, + 0.09056324481300634, + 0.08660473329933704, + 0.08515992711151339, + 0.08358545409687293, + 0.08249965653344751, + 0.08519215440832041, + 0.08217771135284135, + 0.08209042393368436 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.003767186209488777, + 0.0030674240877703748, + 0.0028675942471113798, + 0.0027543867477319494, + 0.0027067201603555215, + 0.002650021515729101, + 0.002624100405593377, + 0.0027113321562234594, + 0.002617328412975119, + 0.002601935845993964 + ], + "val_logit_kl": [ + 0.32186438839566106, + 0.3219214724050983, + 0.24033480349154648, + 0.20962458871464087, + 0.22487169409890925, + 0.20096812819023666, + 0.19121047352584097, + 0.2393866624295549, + 0.19004631119321985, + 0.19876787712503866 + ], + "val_kl_contribution": [ + 0.003218643812694143, + 0.003219214650755346, + 0.002403347981931419, + 0.002096245839181286, + 0.0022487168889549873, + 0.0020096812358551086, + 0.0019121046914511285, + 0.002393866570213465, + 0.0019004630685057397, + 0.0019876787269919364 + ] +} \ No newline at end of file diff --git a/pythia410m_layer16_exp4/weights.pt b/pythia410m_layer16_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..5dab007f3836751271ed9917f1a82820589bbe82 --- /dev/null +++ b/pythia410m_layer16_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b643fa696e34bdbb77a9a4bd9dad0e20373dba13ac300673d364407885ffdb43 +size 33598101 diff --git a/pythia410m_layer20_exp4/config.json b/pythia410m_layer20_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..3fde0b6dce6902d129de7c875156f353f9b3a4ad --- /dev/null +++ b/pythia410m_layer20_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "20", "hook_name": "hook_resid_post", "hook_spec": "blocks.20.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer20_exp4/history.json b/pythia410m_layer20_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..b9b699ae843f7926aad59ae74667ef05127bc542 --- /dev/null +++ b/pythia410m_layer20_exp4/history.json @@ -0,0 +1,222 @@ +{ + "loss": [ + 0.16211504342132493, + 0.10067129953884386, + 0.09280351666689676, + 0.08699335354446229, + 0.08336949619057228, + 0.08215251792654964, + 0.08116990813187673, + 0.08045406652759834, + 0.08055890680032152 + ], + "recon_loss": [ + 0.15564783398880067, + 0.09632656519755911, + 0.088733848639867, + 0.08320554298688464, + 0.07976921046399486, + 0.07864626789610907, + 0.07771860353459356, + 0.07706233202657255, + 0.07715556492941106 + ], + "l1_loss": [ + 0.07019442083067286, + 0.05875471460306764, + 0.05630684493890475, + 0.05551435578844346, + 0.05522927838443574, + 0.054897712533465315, + 0.05467336115591269, + 0.054394984066810884, + 0.053900332916899836 + ], + "sparsity": [ + 127.99681294505262, + 127.9055342042511, + 127.69139970062865, + 127.68776041020499, + 127.68349952490057, + 127.6114986890457, + 127.64406834929478, + 127.68963794050545, + 127.65165958993015 + ], + "recon_contribution": [ + 0.15564783398880067, + 0.09632656519755911, + 0.088733848639867, + 0.08320554298688464, + 0.07976921046399486, + 0.07864626789610907, + 0.07771860353459356, + 0.07706233202657255, + 0.07715556492941106 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.003619418565862932, + 0.0031015758182703027, + 0.0029331032825763185, + 0.002770557122376026, + 0.0026508612855516726, + 0.002613636069342041, + 0.0025777324663133726, + 0.0025438382175684265, + 0.002547697336522832 + ], + "logit_kl": [ + 0.2847791214840328, + 0.12431586245345307, + 0.1136564812107441, + 0.10172533851282539, + 0.0949424416005963, + 0.08926140512094687, + 0.08735721425225995, + 0.08478963147754191, + 0.08556444501967428 + ], + "kl_contribution": [ + 0.0028477911476770326, + 0.0012431585956815953, + 0.001136564786498183, + 0.001017253363524866, + 0.000949424394834643, + 0.0008926140309661963, + 0.0008735721221158493, + 0.0008478962964973156, + 0.0008556444317649331 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.2301682496696426, + 0.20511982654276786, + 0.19695799923360158, + 0.18786643335984166, + 0.18302077987061638, + 0.18684740998070926, + 0.17843776434890143, + 0.19148184931153658, + 0.17577280922336289 + ], + "val_recon_loss": [ + 0.21850402330956856, + 0.1953204816401086, + 0.18739983046775188, + 0.17901436925130723, + 0.17441177826166587, + 0.17760054877654957, + 0.17005954135458126, + 0.17914778125497333, + 0.16759699058703323 + ], + "val_l1_loss": [ + 0.10386819195261277, + 0.08928536385575353, + 0.08236716298224817, + 0.07651199199858442, + 0.07403400410702964, + 0.07176928357582968, + 0.07048508180800431, + 0.06899321687723874, + 0.06810225372874211 + ], + "val_sparsity": [ + 127.9975695958846, + 127.98939837915472, + 127.97696320439728, + 127.96649647235654, + 127.9721652407475, + 127.97337379784868, + 127.9734211548628, + 127.96097953272971, + 127.98348645457675 + ], + "val_recon_contribution": [ + 0.21850402330956856, + 0.1953204816401086, + 0.18739983046775188, + 0.17901436925130723, + 0.17441177826166587, + 0.17760054877654957, + 0.17005954135458126, + 0.17914778125497333, + 0.16759699058703323 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0069029904664602, + 0.0062348846392050175, + 0.006025284951339488, + 0.005700396759133415, + 0.005613630053175546, + 0.005654315462770505, + 0.005434500712242107, + 0.005724624889534219, + 0.005348385494034814 + ], + "val_logit_kl": [ + 0.4761236185027707, + 0.35644603614060244, + 0.3532883652411012, + 0.3151667239698753, + 0.2995371618432952, + 0.3592545676846116, + 0.2943722312150355, + 0.6609443491667523, + 0.2827433166836728 + ], + "val_kl_contribution": [ + 0.004761236077995428, + 0.003564460284901958, + 0.0035328835714714215, + 0.0031516671700970943, + 0.0029953715524163142, + 0.003592545598765868, + 0.002943722245873488, + 0.0066094433456982595, + 0.0028274331021233496 + ] +} \ No newline at end of file diff --git a/pythia410m_layer20_exp4/weights.pt b/pythia410m_layer20_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..fb71dc02e81a36986bb27dd0706aeb65a11f9367 --- /dev/null +++ b/pythia410m_layer20_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ab7f4748676278f9c95c91c807c83980ef9ff77e2521c75b75a33d062556064 +size 33598101 diff --git a/pythia410m_layer23_exp4/config.json b/pythia410m_layer23_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..070418feb886825fabd620e177a692abaa208edb --- /dev/null +++ b/pythia410m_layer23_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "23", "hook_name": "hook_resid_post", "hook_spec": "blocks.23.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer23_exp4/history.json b/pythia410m_layer23_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..553e8ec4686d5065e77fbb589b71553e98ca2e07 --- /dev/null +++ b/pythia410m_layer23_exp4/history.json @@ -0,0 +1,222 @@ +{ + "loss": [ + 0.19471221902794772, + 0.1359062709891239, + 0.1277088505231155, + 0.12214199725044941, + 0.1202180515805736, + 0.11676322992978966, + 0.115552180465277, + 0.11407808340687015, + 0.11555235020938023 + ], + "recon_loss": [ + 0.19232082875029685, + 0.13457462265614265, + 0.12647040038050142, + 0.12096263629007416, + 0.11905944195428961, + 0.11564715564785062, + 0.11444143741687392, + 0.11299694314457057, + 0.11445592510464944 + ], + "l1_loss": [ + 0.055780474241897325, + 0.05456285348868575, + 0.05571047466133337, + 0.05619010674731701, + 0.056660164546512216, + 0.056776592727689694, + 0.056775683729813016, + 0.05716668963668946, + 0.05740164942741935 + ], + "sparsity": [ + 127.99998547161123, + 127.99983322988189, + 127.99986248050973, + 127.99983973771388, + 127.99978587155766, + 127.9997793706489, + 127.99977200087531, + 127.99970554483347, + 127.99962357038162 + ], + "recon_contribution": [ + 0.19232082875029685, + 0.13457462265614265, + 0.12647040038050142, + 0.12096263629007416, + 0.11905944195428961, + 0.11564715564785062, + 0.11444143741687392, + 0.11299694314457057, + 0.11445592510464944 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "logit_kl": [ + 0.2391390282738474, + 0.13316483657803732, + 0.12384502456427135, + 0.11793610287539512, + 0.11586095334475222, + 0.1116074294151025, + 0.11107431043923145, + 0.10811403353433184, + 0.10964251046631816 + ], + "kl_contribution": [ + 0.002391390229333737, + 0.0013316483353383475, + 0.0012384502178298042, + 0.0011793610017333423, + 0.0011586095068843345, + 0.001116074269209412, + 0.0011107430793556256, + 0.001081140311065653, + 0.0010964250803543526 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.316197270421698, + 0.2933344059487951, + 0.2857086572056464, + 0.2835407973804944, + 0.277085559516265, + 0.2770049967166241, + 0.27574585438763105, + 0.2738202984551851, + 0.2707638205966317 + ], + "val_recon_loss": [ + 0.31154836239678274, + 0.289137319403987, + 0.2816178682296182, + 0.2795479790337624, + 0.2732222687210998, + 0.2730984420010795, + 0.2718467891338141, + 0.2699995644996514, + 0.2669640527543075 + ], + "val_l1_loss": [ + 0.06124183855974528, + 0.06370097175849833, + 0.0660762913653738, + 0.06600362401978636, + 0.06667904545303802, + 0.06654060923615839, + 0.06765059681811532, + 0.06747716691671637, + 0.06754671838955412 + ], + "val_sparsity": [ + 127.99999902596099, + 127.99998284512787, + 127.99997265411474, + 127.99988684509083, + 127.99996130188072, + 127.99994086092681, + 127.99990802090691, + 127.99997085855529, + 127.99995167587953 + ], + "val_recon_contribution": [ + 0.31154836239678274, + 0.289137319403987, + 0.2816178682296182, + 0.2795479790337624, + 0.2732222687210998, + 0.2730984420010795, + 0.2718467891338141, + 0.2699995644996514, + 0.2669640527543075 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.46489082612283444, + 0.41970864466743973, + 0.4090789091696744, + 0.3992818322054693, + 0.3863290916419148, + 0.3906554739206339, + 0.3899065362410565, + 0.38207343585284825, + 0.379976795738337 + ], + "val_kl_contribution": [ + 0.0046489081529668555, + 0.004197086352757223, + 0.004090789005092419, + 0.0039928182316654156, + 0.003863290828857478, + 0.003906554652470842, + 0.0038990652727258087, + 0.0038207342707171615, + 0.003799767870171844 + ] +} \ No newline at end of file diff --git a/pythia410m_layer23_exp4/weights.pt b/pythia410m_layer23_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..7be9aecc80f96ed471c8206626d1ae4f9a6a81ea --- /dev/null +++ b/pythia410m_layer23_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6448973ed9c9c14ba995c7aea4ac5d33682a57aba29202332d3846c72fd55572 +size 33598101 diff --git a/pythia410m_layer4_exp4/config.json b/pythia410m_layer4_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..1cf6e6f2c732afe7cbd8ba3a8007ff266c9b845b --- /dev/null +++ b/pythia410m_layer4_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "4", "hook_name": "hook_resid_post", "hook_spec": "blocks.4.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer4_exp4/history.json b/pythia410m_layer4_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..8872a75abdb7dd8cc50396fa5c5fa55f3aacfadc --- /dev/null +++ b/pythia410m_layer4_exp4/history.json @@ -0,0 +1,142 @@ +{ + "loss": [ + 0.029894512988028612, + 0.018774982353548998, + 0.01757451586409247, + 0.016827763248608532, + 0.01671755956206978 + ], + "recon_loss": [ + 0.028509802402828747, + 0.017858383666847354, + 0.016714493958947134, + 0.016002113766082564, + 0.015895764886641607 + ], + "l1_loss": [ + 0.030895500990668626, + 0.02851335875632515, + 0.027246419608275223, + 0.02655725095407554, + 0.026199420892925908 + ], + "sparsity": [ + 127.86990778762069, + 127.20474829215104, + 126.57427332924844, + 126.12415448377439, + 126.17486108409514 + ], + "recon_contribution": [ + 0.028509802402828747, + 0.017858383666847354, + 0.016714493958947134, + 0.016002113766082564, + 0.015895764886641607 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0006074781647182751, + 0.0005631423298345148, + 0.0005285829312475791, + 0.000506121144354717, + 0.0005024663806066721 + ], + "logit_kl": [ + 0.07772324425328908, + 0.03534563696527424, + 0.03314389820006117, + 0.03195283477390675, + 0.03193282936116555 + ], + "kl_contribution": [ + 0.0007772324256051311, + 0.0003534563615086878, + 0.0003314389744292581, + 0.0003195283404557151, + 0.0003193282865419601 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.05225575266271383, + 0.04949509506847968, + 0.04847414334131568, + 0.04836168864751231, + 0.04753865694363975 + ], + "val_recon_loss": [ + 0.04925031143315526, + 0.04667543858985354, + 0.04570288170486188, + 0.04559107506220338, + 0.044836636487968345 + ], + "val_l1_loss": [ + 0.030398870868670134, + 0.029119303775803167, + 0.028552788553902475, + 0.028373856366574357, + 0.028430795825100497 + ], + "val_sparsity": [ + 127.99124588504914, + 127.91639768746916, + 127.88443639288161, + 127.73844488508752, + 127.8981660100234 + ], + "val_recon_contribution": [ + 0.04925031143315526, + 0.04667543858985354, + 0.04570288170486188, + 0.04559107506220338, + 0.044836636487968345 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0015425308895145443, + 0.0014635327424499028, + 0.0014332294819043186, + 0.0014288228828355674, + 0.0014068659569972705 + ], + "val_logit_kl": [ + 0.14629103523297832, + 0.1356123878651454, + 0.133803220682425, + 0.1341790673617055, + 0.1295154520311731 + ], + "val_kl_contribution": [ + 0.0014629103198525002, + 0.0013561238491139088, + 0.0013380321782384967, + 0.0013417906447417182, + 0.0012951544919965203 + ] +} \ No newline at end of file diff --git a/pythia410m_layer4_exp4/weights.pt b/pythia410m_layer4_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..f1a5aecd6734a51938c7302c757726b19913baec --- /dev/null +++ b/pythia410m_layer4_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:708b2b5dea22194358bde12466e8875c4e594bd8ea14d18801e4d28e97221cc8 +size 33598101 diff --git a/pythia410m_layer8_exp4/config.json b/pythia410m_layer8_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..97e61cec441c1f6fdecacfe7754b2fc633db922e --- /dev/null +++ b/pythia410m_layer8_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 1024, "d_sae": 4096, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "8", "hook_name": "hook_resid_post", "hook_spec": "blocks.8.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia410m_layer8_exp4/history.json b/pythia410m_layer8_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..046393d23b4972052bf313eae03acfac7112f239 --- /dev/null +++ b/pythia410m_layer8_exp4/history.json @@ -0,0 +1,182 @@ +{ + "loss": [ + 0.06778999520279805, + 0.039218354890352886, + 0.035718238334883065, + 0.03308342497528762, + 0.031953453523666346, + 0.030800824367547722, + 0.029794476944074277 + ], + "recon_loss": [ + 0.06487649836714385, + 0.03731946416576131, + 0.033973340638499105, + 0.03149048615197994, + 0.030414512773093424, + 0.029341977546431003, + 0.02840521304019661 + ], + "l1_loss": [ + 0.04243372238584255, + 0.037460498501649786, + 0.03468995287629037, + 0.033307530635802204, + 0.03245061089600793, + 0.03202774247556871, + 0.03166854607292913 + ], + "sparsity": [ + 127.99999613684015, + 127.99979148283195, + 127.99808386233238, + 127.99375951874278, + 127.98073970773908, + 127.97515894064237, + 127.95779129900647 + ], + "recon_contribution": [ + 0.06487649836714385, + 0.03731946416576131, + 0.033973340638499105, + 0.03149048615197994, + 0.030414512773093424, + 0.029341977546431003, + 0.02840521304019661 + ], + "l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "aux_loss": [ + 0.00128651277245379, + 0.0011709438835739957, + 0.0010682826464611845, + 0.0009930854126442977, + 0.0009616236588401093, + 0.0009283572463505688, + 0.0008989799316887348 + ], + "logit_kl": [ + 0.16269841401697477, + 0.07279468437541824, + 0.06766150737039325, + 0.059985338663344295, + 0.057731709948414686, + 0.05304895968222061, + 0.04902839502607034 + ], + "kl_contribution": [ + 0.0016269841030927352, + 0.0007279468274758448, + 0.0006766150578721883, + 0.0005998533735078002, + 0.0005773170866950997, + 0.0005304895850528495, + 0.0004902839390806861 + ], + "dead_features": [ + 0, + 0, + 0, + 0, + 0, + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_loss": [ + 0.10301138475869039, + 0.09221573064153653, + 0.07529819821316862, + 0.07469267990171016, + 0.08256304211938235, + 0.07100475511206031, + 0.07067192238296421 + ], + "val_recon_loss": [ + 0.09424476893022081, + 0.0869976317697035, + 0.07134533901087456, + 0.0706223731793663, + 0.07826717641849455, + 0.06723475776213318, + 0.0669424738176421 + ], + "val_l1_loss": [ + 0.08854262023209605, + 0.06788539300854235, + 0.05627321405865181, + 0.05003325807447734, + 0.04645337848189959, + 0.043372284923119474, + 0.04172941518677521 + ], + "val_sparsity": [ + 127.99999902596099, + 127.99999382993082, + 127.99993775856056, + 127.99855484501008, + 127.99967206082307, + 127.99946186944389, + 127.99578929868193 + ], + "val_recon_contribution": [ + 0.09424476893022081, + 0.0869976317697035, + 0.07134533901087456, + 0.0706223731793663, + 0.07826717641849455, + 0.06723475776213318, + 0.0669424738176421 + ], + "val_l1_contribution": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.002948156130503946, + 0.002725725934698142, + 0.0022369817911441193, + 0.0022149345621099354, + 0.0024611907477722485, + 0.0021109028090691125, + 0.002104920502199873 + ], + "val_logit_kl": [ + 0.5818460128460078, + 0.24923730375535486, + 0.17158774770745522, + 0.1855372296530298, + 0.1834674852306547, + 0.16590945168936974, + 0.16245281352563806 + ], + "val_kl_contribution": [ + 0.00581845999839227, + 0.002492372982542681, + 0.001715877436233984, + 0.0018553722541965884, + 0.0018346748123434574, + 0.0016590944811297045, + 0.0016245280985334556 + ] +} \ No newline at end of file diff --git a/pythia410m_layer8_exp4/weights.pt b/pythia410m_layer8_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..c11f40fb15cd879e4a374cd4b8984f09e4d1125a --- /dev/null +++ b/pythia410m_layer8_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:027e390c5753cccabba0e9d57a4cf31418301d06046d203a6f58763a68bb7103 +size 33598101 diff --git a/pythia70m_layer0_exp4/config.json b/pythia70m_layer0_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b8eca7edebddde335e942bf1dc0d238acd0904d4 --- /dev/null +++ b/pythia70m_layer0_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 512, "d_sae": 2048, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "0", "hook_name": "hook_resid_post", "hook_spec": "blocks.0.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia70m_layer0_exp4/history.json b/pythia70m_layer0_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..af754a1550924629a562ef0f26f82bd6e1a5e935 --- /dev/null +++ b/pythia70m_layer0_exp4/history.json @@ -0,0 +1,82 @@ +{ + "loss": [ + 0.009120507270161962, + 0.004647127874094162 + ], + "recon_loss": [ + 0.007533900376777765, + 0.0037761096149165783 + ], + "l1_loss": [ + 0.021236359151908468, + 0.020717918377592754 + ], + "sparsity": [ + 127.99993658325417, + 127.77261757547755 + ], + "recon_contribution": [ + 0.007533900376777765, + 0.0037761096149165783 + ], + "l1_contribution": [ + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0 + ], + "logit_kl": [ + 0.15866069382431827, + 0.08710182806689921 + ], + "kl_contribution": [ + 0.0015866069037897407, + 0.0008710182613960235 + ], + "dead_features": [ + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0 + ], + "val_loss": [ + 0.011327041563434994, + 0.009696548539666386 + ], + "val_recon_loss": [ + 0.009243751145843467, + 0.007814783182123728 + ], + "val_l1_loss": [ + 0.02203361001491655, + 0.022379753541042328 + ], + "val_sparsity": [ + 127.99998241982387, + 127.97187148950839 + ], + "val_recon_contribution": [ + 0.009243751145843467, + 0.007814783182123728 + ], + "val_l1_contribution": [ + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.2083290464682107, + 0.1881765398339831 + ], + "val_kl_contribution": [ + 0.002083290415635412, + 0.0018817653529431425 + ] +} \ No newline at end of file diff --git a/pythia70m_layer0_exp4/weights.pt b/pythia70m_layer0_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..ced4733dee9a7072d0eda4cf1d679e22cd091b53 --- /dev/null +++ b/pythia70m_layer0_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:370906c774a822b712282410e8306f97966ed7e238425385f52476843a8ab647 +size 8411797 diff --git a/pythia70m_layer2_exp4/config.json b/pythia70m_layer2_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..1bba5f42799c3c4a55e13cd62192e23264b62623 --- /dev/null +++ b/pythia70m_layer2_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 512, "d_sae": 2048, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "2", "hook_name": "hook_resid_post", "hook_spec": "blocks.2.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia70m_layer2_exp4/history.json b/pythia70m_layer2_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..3d794f24a4c7ced32cae1222c090de7d568f3bad --- /dev/null +++ b/pythia70m_layer2_exp4/history.json @@ -0,0 +1,82 @@ +{ + "loss": [ + 0.022089810386857613, + 0.011305966657953857 + ], + "recon_loss": [ + 0.020042146756915617, + 0.010267486703363532 + ], + "l1_loss": [ + 0.04374750378180523, + 0.05297717931268572 + ], + "sparsity": [ + 127.9999990969288, + 127.99946044091465 + ], + "recon_contribution": [ + 0.020042146756915617, + 0.010267486703363532 + ], + "l1_contribution": [ + 0.0, + 0.0 + ], + "aux_loss": [ + 0.0, + 0.0 + ], + "logit_kl": [ + 0.20476636848836935, + 0.10384799798088921 + ], + "kl_contribution": [ + 0.0020476636394056915, + 0.0010384799558178326 + ], + "dead_features": [ + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0 + ], + "val_loss": [ + 0.026872203112710026, + 0.022523965515785463 + ], + "val_recon_loss": [ + 0.02400330499312304, + 0.020051311383794564 + ], + "val_l1_loss": [ + 0.07073753580054866, + 0.06914176224605843 + ], + "val_sparsity": [ + 127.99999902227255, + 127.99999902227255 + ], + "val_recon_contribution": [ + 0.02400330499312304, + 0.020051311383794564 + ], + "val_l1_contribution": [ + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.0, + 0.0 + ], + "val_logit_kl": [ + 0.2868898172076251, + 0.24726541762398307 + ], + "val_kl_contribution": [ + 0.002868898107604223, + 0.0024726541196113555 + ] +} \ No newline at end of file diff --git a/pythia70m_layer2_exp4/weights.pt b/pythia70m_layer2_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..2a37cd8ca400c15315ea0e468d208e004b6674d0 --- /dev/null +++ b/pythia70m_layer2_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:270b058168ebd748f5c731d61b37f1f465661a3e0d4f7a5440619113d8a8bdd1 +size 8411797 diff --git a/pythia70m_layer5_exp4/config.json b/pythia70m_layer5_exp4/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8d8a14c14393ee9332f24e45edb77c8cdafbb392 --- /dev/null +++ b/pythia70m_layer5_exp4/config.json @@ -0,0 +1 @@ +{"d_in": 512, "d_sae": 2048, "l1_coefficient": 0.0, "dtype": "float32", "device": "cuda", "hook_layer": "5", "hook_name": "hook_resid_post", "hook_spec": "blocks.5.hook_resid_post", "top_k": 128, "top_k_aux": 64, "n_batches_to_dead": 200, "aux_penalty": 0.03125} \ No newline at end of file diff --git a/pythia70m_layer5_exp4/history.json b/pythia70m_layer5_exp4/history.json new file mode 100644 index 0000000000000000000000000000000000000000..3cbb4dcbee9a7ed2d0202a913ed67159967c8676 --- /dev/null +++ b/pythia70m_layer5_exp4/history.json @@ -0,0 +1,82 @@ +{ + "loss": [ + 0.15055845827607947, + 0.09721381716707386 + ], + "recon_loss": [ + 0.13905752800445592, + 0.09052910425357129 + ], + "l1_loss": [ + 0.15998336581509906, + 0.14880641750626253 + ], + "sparsity": [ + 127.99999728263398, + 127.99935442998924 + ], + "recon_contribution": [ + 0.13905752800445592, + 0.09052910425357129 + ], + "l1_contribution": [ + 0.0, + 0.0 + ], + "aux_loss": [ + 0.003473982798800163, + 0.0028626951294466888 + ], + "logit_kl": [ + 0.8026947751877908, + 0.3822017896298434 + ], + "kl_contribution": [ + 0.008026947576509106, + 0.003822017813054084 + ], + "dead_features": [ + 0, + 0 + ], + "dead_feature_percentage": [ + 0.0, + 0.0 + ], + "val_loss": [ + 0.1655603101961941, + 0.14054889555703992 + ], + "val_recon_loss": [ + 0.1534292830686605, + 0.13083974581917596 + ], + "val_l1_loss": [ + 0.16197139682118755, + 0.142402270438995 + ], + "val_sparsity": [ + 127.999997025888, + 127.99999902596099 + ], + "val_recon_contribution": [ + 0.1534292830686605, + 0.13083974581917596 + ], + "val_l1_contribution": [ + 0.0, + 0.0 + ], + "val_aux_loss": [ + 0.004870677601587516, + 0.00413685759079043 + ], + "val_logit_kl": [ + 0.7260349912637475, + 0.5572292297123942 + ], + "val_kl_contribution": [ + 0.007260349753275954, + 0.00557229217410135 + ] +} \ No newline at end of file diff --git a/pythia70m_layer5_exp4/weights.pt b/pythia70m_layer5_exp4/weights.pt new file mode 100644 index 0000000000000000000000000000000000000000..a2e4d4cfcc48e97c6ed2f72bfbcee0ad5b7d9e23 --- /dev/null +++ b/pythia70m_layer5_exp4/weights.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89dcf8c010e43b7e717c83040df6d57178866bf320db345fc84f926b9c54f527 +size 8411797