Upload add_sub_sorl_v1_abs10_K1_25K

add_sub_sorl_v1_abs10_K1_25K/metrics.json

@@ -73,509 +73,509 @@
       3869
     ],
     "loss": [
-      8.440483093261719,
-      3.670170307159424,
-      3.2093284130096436,
-      3.1004648208618164,
-      3.070091724395752,
-      3.0389091968536377,
-      2.5894856452941895,
-      0.7675281763076782,
-      -3.0276317596435547,
-      -4.935643196105957,
-      -4.584414958953857,
-      -2.691551923751831,
-      -1.9160187244415283,
-      -1.2295455932617188,
-      -1.3427369594573975,
-      -1.812907099723816,
-      -1.9342525005340576,
-      -1.293236255645752,
-      -1.3251506090164185,
-      -1.4479036331176758,
-      -1.5105352401733398,
-      -2.014406204223633,
-      -1.9828301668167114,
-      -1.7031790018081665,
-      -2.5977749824523926,
-      -2.3531222343444824,
-      -2.0115716457366943,
-      -1.536534070968628,
-      -1.4918975830078125,
-      -1.0497915744781494,
-      -1.3855948448181152,
-      -1.9143116474151611,
-      -1.2018264532089233,
-      -0.9971982836723328,
-      -1.4220505952835083,
-      -0.9136620759963989,
-      -0.9068200588226318,
-      -1.1527667045593262,
-      -0.6265409588813782,
-      -0.8574066162109375,
-      -0.42161595821380615,
-      -0.434919536113739,
-      -0.4632154107093811,
-      -0.5082250237464905,
-      -0.5800994634628296,
-      -0.789692759513855,
-      -0.2702712416648865,
-      -0.4394497275352478,
-      -0.5490878820419312,
-      -0.5185199975967407,
-      -0.32845696806907654,
-      -0.2897554337978363,
-      -0.39504489302635193,
-      -0.27258846163749695,
-      -0.25301221013069153,
-      -0.08269159495830536,
-      -0.18398964405059814,
-      -0.23646067082881927,
-      -0.10407372564077377,
-      -0.09299999475479126,
-      -0.08140064775943756,
-      -0.09688820689916611,
-      -0.05483222380280495,
-      -0.03655329346656799,
-      -0.05594222992658615,
-      -0.008253187872469425,
-      -0.09734220057725906,
-      -0.05176986753940582,
-      -0.05214906856417656,
-      -0.07431081682443619
+      2.775237560272217,
+      7.547610282897949,
+      4.404772758483887,
+      3.531778335571289,
+      3.0909929275512695,
+      3.0005178451538086,
+      2.930567741394043,
+      2.380660057067871,
+      1.8402857780456543,
+      1.5253196954727173,
+      0.6456812620162964,
+      -0.922775387763977,
+      -5.972268104553223,
+      -7.329289436340332,
+      -3.549975872039795,
+      -2.491389274597168,
+      -0.8877972364425659,
+      -0.7063808441162109,
+      -1.138548493385315,
+      -0.44369208812713623,
+      -0.18519365787506104,
+      0.001989603042602539,
+      -0.2683281898498535,
+      0.038404107093811035,
+      0.15909254550933838,
+      0.3572588562965393,
+      -0.12331002950668335,
+      0.23269200325012207,
+      -0.06346249580383301,
+      -0.600041389465332,
+      -0.17923784255981445,
+      -0.3249526619911194,
+      0.10653692483901978,
+      -0.6031087636947632,
+      -0.2606835961341858,
+      -0.31056416034698486,
+      -0.71147620677948,
+      -0.4193435311317444,
+      -0.8142799139022827,
+      -0.7582582235336304,
+      -0.38260582089424133,
+      -0.4925423562526703,
+      -0.6277319192886353,
+      -0.5789560675621033,
+      -0.5000548958778381,
+      -0.24027329683303833,
+      -0.5376996994018555,
+      -0.7423339486122131,
+      -0.6099227070808411,
+      -0.734022855758667,
+      -0.5901182293891907,
+      -0.4833343029022217,
+      -0.50217604637146,
+      -0.4287641942501068,
+      -0.42581605911254883,
+      -0.32260385155677795,
+      -0.4461747407913208,
+      -0.4050253629684448,
+      -0.6488711833953857,
+      -0.3880271315574646,
+      -0.5698164105415344,
+      -0.6381738781929016,
+      -0.5308898091316223,
+      -0.3668437600135803,
+      -0.5520040988922119,
+      -0.463625431060791,
+      -0.6266764402389526,
+      -0.4801936149597168,
+      -0.22993049025535583,
+      -0.4441494047641754
     ],
     "base_loss": [
-      6.488893985748291,
-      2.3548998832702637,
-      1.9527490139007568,
-      1.8523368835449219,
-      1.8802011013031006,
-      1.8292043209075928,
-      1.8652232885360718,
-      1.815848469734192,
-      1.848860740661621,
-      1.741455316543579,
-      1.3914107084274292,
-      1.068152666091919,
-      0.9446887969970703,
-      0.7281270623207092,
-      0.599748432636261,
-      0.5804276466369629,
-      0.5249930024147034,
-      0.454721599817276,
-      0.3896333873271942,
-      0.32255008816719055,
-      0.34916767477989197,
-      0.32530543208122253,
-      0.2995271384716034,
-      0.2648261487483978,
-      0.3412800133228302,
-      0.28785938024520874,
-      0.24230030179023743,
-      0.19102659821510315,
-      0.17416465282440186,
-      0.1800290048122406,
-      0.1682046502828598,
-      0.24820947647094727,
-      0.14416953921318054,
-      0.11789846420288086,
-      0.16254489123821259,
-      0.11386360228061676,
-      0.12063971906900406,
-      0.13949914276599884,
-      0.07232778519392014,
-      0.10645522177219391,
-      0.04950258880853653,
-      0.05214657261967659,
-      0.0541020967066288,
-      0.06009664013981819,
-      0.07329648733139038,
-      0.09814182668924332,
-      0.031828977167606354,
-      0.05065033212304115,
-      0.06576144695281982,
-      0.06058066338300705,
-      0.037494949996471405,
-      0.03307747095823288,
-      0.04483410716056824,
-      0.03311517834663391,
-      0.030032988637685776,
-      0.00985543243587017,
-      0.02162509225308895,
-      0.02781860902905464,
-      0.012631254270672798,
-      0.011143012903630733,
-      0.009497015736997128,
-      0.01135705504566431,
-      0.008042754605412483,
-      0.005164582747966051,
-      0.006489872932434082,
-      0.0014688527444377542,
-      0.011420228518545628,
-      0.006360507570207119,
-      0.006299526430666447,
-      0.008574886247515678
+      9.28781795501709,
+      6.408207416534424,
+      4.156429767608643,
+      2.373605966567993,
+      2.0184969902038574,
+      1.929623007774353,
+      1.8580621480941772,
+      1.8653138875961304,
+      1.8420647382736206,
+      1.7876325845718384,
+      1.7426172494888306,
+      1.8583365678787231,
+      1.8177855014801025,
+      1.7561644315719604,
+      1.1391898393630981,
+      0.8618606925010681,
+      0.6386224627494812,
+      0.6016677021980286,
+      0.5611599087715149,
+      0.43799519538879395,
+      0.39136165380477905,
+      0.35541364550590515,
+      0.3650291860103607,
+      0.28820252418518066,
+      0.28319448232650757,
+      0.27565693855285645,
+      0.26511654257774353,
+      0.17914986610412598,
+      0.23786573112010956,
+      0.3100714087486267,
+      0.20881570875644684,
+      0.2222459614276886,
+      0.1805049628019333,
+      0.21834571659564972,
+      0.17398054897785187,
+      0.17169272899627686,
+      0.214468851685524,
+      0.16018731892108917,
+      0.21521098911762238,
+      0.1976483017206192,
+      0.16451016068458557,
+      0.15567553043365479,
+      0.1631745547056198,
+      0.17332789301872253,
+      0.16741959750652313,
+      0.14030645787715912,
+      0.16195133328437805,
+      0.19442276656627655,
+      0.23462989926338196,
+      0.1750485599040985,
+      0.15816131234169006,
+      0.1345469057559967,
+      0.11074493825435638,
+      0.09530355036258698,
+      0.12718015909194946,
+      0.1237618550658226,
+      0.15354827046394348,
+      0.0893278494477272,
+      0.1284264326095581,
+      0.0889713391661644,
+      0.10795746743679047,
+      0.11995094269514084,
+      0.11075253039598465,
+      0.09607528150081635,
+      0.0994424819946289,
+      0.13344082236289978,
+      0.12076304107904434,
+      0.10045450925827026,
+      0.07217186689376831,
+      0.08645479381084442
     ],
     "info_loss": [
-      -0.2627429962158203,
-      -0.0675959587097168,
-      -0.06391513347625732,
-      -0.06318938732147217,
-      -0.068672776222229,
-      -0.06626904010772705,
-      -0.11490964889526367,
-      -0.29147613048553467,
-      -0.6731091737747192,
-      -0.8447501063346863,
-      -0.7606219053268433,
-      -0.516476035118103,
-      -0.41259610652923584,
-      -0.31304433941841125,
-      -0.2955251634120941,
-      -0.3318001329898834,
-      -0.33346521854400635,
-      -0.25125789642333984,
-      -0.23548583686351776,
-      -0.21368299424648285,
-      -0.20429086685180664,
-      -0.24210147559642792,
-      -0.23295104503631592,
-      -0.19994628429412842,
-      -0.2960708439350128,
-      -0.2660777270793915,
-      -0.22968997061252594,
-      -0.17489448189735413,
-      -0.16804951429367065,
-      -0.12430460005998611,
-      -0.15695549547672272,
-      -0.2182704359292984,
-      -0.13539204001426697,
-      -0.11300598084926605,
-      -0.15940292179584503,
-      -0.10396484285593033,
-      -0.10412997007369995,
-      -0.13012480735778809,
-      -0.07035745680332184,
-      -0.09725693613290787,
-      -0.0479651540517807,
-      -0.04964497312903404,
-      -0.05318129435181618,
-      -0.05739910155534744,
-      -0.06728154420852661,
-      -0.09058966487646103,
-      -0.03128577023744583,
-      -0.04990841820836067,
-      -0.06233469396829605,
-      -0.05823947489261627,
-      -0.03711993992328644,
-      -0.03283752128481865,
-      -0.04459574073553085,
-      -0.031027397140860558,
-      -0.028812875971198082,
-      -0.009709836915135384,
-      -0.021456001326441765,
-      -0.02661510370671749,
-      -0.012435774318873882,
-      -0.011000837199389935,
-      -0.009408645331859589,
-      -0.011241589672863483,
-      -0.006964982487261295,
-      -0.005082710646092892,
-      -0.006405099760740995,
-      -0.0013885911321267486,
-      -0.011316562071442604,
-      -0.006281006615608931,
-      -0.006225514691323042,
-      -0.008512690663337708
+      -1.4931654930114746,
+      -0.3287515640258789,
+      -0.20309758186340332,
+      -0.08040070533752441,
+      -0.08293402194976807,
+      -0.08145439624786377,
+      -0.08084726333618164,
+      -0.13623690605163574,
+      -0.18764793872833252,
+      -0.21354830265045166,
+      -0.2971622943878174,
+      -0.4655829668045044,
+      -0.9671027064323425,
+      -1.0975041389465332,
+      -0.6573877930641174,
+      -0.52280592918396,
+      -0.33613646030426025,
+      -0.30700165033340454,
+      -0.33531850576400757,
+      -0.24739503860473633,
+      -0.20575667917728424,
+      -0.1729479283094406,
+      -0.19086585938930511,
+      -0.1466362029314041,
+      -0.12648341059684753,
+      -0.09660688042640686,
+      -0.13731198012828827,
+      -0.090257927775383,
+      -0.11520253866910934,
+      -0.16884556412696838,
+      -0.1119970828294754,
+      -0.12500318884849548,
+      -0.07407157123088837,
+      -0.1408563256263733,
+      -0.10861283540725708,
+      -0.10317053645849228,
+      -0.1429494023323059,
+      -0.10317841172218323,
+      -0.14690537750720978,
+      -0.13502144813537598,
+      -0.0957949236035347,
+      -0.10858335345983505,
+      -0.11433392018079758,
+      -0.11800597608089447,
+      -0.1043647900223732,
+      -0.07300866395235062,
+      -0.10729879140853882,
+      -0.12777504324913025,
+      -0.1146274283528328,
+      -0.11757340282201767,
+      -0.10103149712085724,
+      -0.0873810350894928,
+      -0.08418776094913483,
+      -0.07348176836967468,
+      -0.08262161910533905,
+      -0.06471286714076996,
+      -0.08114015311002731,
+      -0.07110169529914856,
+      -0.09566681832075119,
+      -0.07004982978105545,
+      -0.08596527576446533,
+      -0.09278135001659393,
+      -0.08236707746982574,
+      -0.06275559961795807,
+      -0.08083701133728027,
+      -0.07647520303726196,
+      -0.08978942036628723,
+      -0.07253386080265045,
+      -0.04718723148107529,
+      -0.06880763173103333
     ],
     "abs_loss": [
-      2.062779664993286,
-      1.8660074472427368,
-      1.8702964782714844,
-      1.8314086198806763,
-      1.8298479318618774,
-      1.8219594955444336,
-      1.854802131652832,
-      1.825403094291687,
-      1.8147176504135132,
-      1.718223214149475,
-      1.5588526725769043,
-      1.2659032344818115,
-      1.0050292015075684,
-      0.8381269574165344,
-      0.7240123748779297,
-      0.6084690690040588,
-      0.56196129322052,
-      0.43736353516578674,
-      0.4195970594882965,
-      0.3356196880340576,
-      0.3217657208442688,
-      0.18604372441768646,
-      0.1829766184091568,
-      0.15309858322143555,
-      0.12250169366598129,
-      0.09407112747430801,
-      0.11580777168273926,
-      0.09737056493759155,
-      0.06344643235206604,
-      0.06409666687250137,
-      0.06125688925385475,
-      0.04178151860833168,
-      0.03453303501009941,
-      0.045983314514160156,
-      0.0417625792324543,
-      0.037050385028123856,
-      0.03088521957397461,
-      0.01876666396856308,
-      0.02520829439163208,
-      0.032921046018600464,
-      0.02231394685804844,
-      0.013566468842327595,
-      0.01799575798213482,
-      0.020950788632035255,
-      0.025772780179977417,
-      0.014466878958046436,
-      0.010489404201507568,
-      0.007517005782574415,
-      0.012727208435535431,
-      0.010788698680698872,
-      0.012184095568954945,
-      0.009171529673039913,
-      0.003999281208962202,
-      0.00763775035738945,
-      0.008048650808632374,
-      0.007546226028352976,
-      0.04280756413936615,
-      0.004729445558041334,
-      0.0336616076529026,
-      0.003961015492677689,
-      0.003912051673978567,
-      0.002861393615603447,
-      0.006178874522447586,
-      0.002068899804726243,
-      0.004654192831367254,
-      0.004825535695999861,
-      0.0032716484274715185,
-      0.004482524935156107,
-      0.004428912419825792,
-      0.004628469236195087
+      2.287022590637207,
+      2.0916876792907715,
+      1.904157280921936,
+      1.865134596824646,
+      1.8645482063293457,
+      1.8384100198745728,
+      1.8436139822006226,
+      1.8555344343185425,
+      1.8272294998168945,
+      1.8342450857162476,
+      1.8453730344772339,
+      1.838832974433899,
+      1.8428555727005005,
+      1.8253898620605469,
+      1.7948323488235474,
+      1.7442383766174316,
+      1.575785517692566,
+      1.3669052124023438,
+      1.1249511241912842,
+      0.8671944737434387,
+      0.7276226878166199,
+      0.617413341999054,
+      0.5250348448753357,
+      0.44361987709999084,
+      0.42031577229499817,
+      0.41864728927612305,
+      0.43554285168647766,
+      0.3743334114551544,
+      0.34732961654663086,
+      0.33214130997657776,
+      0.3264453709125519,
+      0.31868138909339905,
+      0.3359517753124237,
+      0.2563074827194214,
+      0.2655089795589447,
+      0.24500541388988495,
+      0.28663334250450134,
+      0.2430928349494934,
+      0.24621863663196564,
+      0.22454464435577393,
+      0.22838687896728516,
+      0.23042821884155273,
+      0.2049625664949417,
+      0.21573372185230255,
+      0.21731531620025635,
+      0.21410906314849854,
+      0.1734648495912552,
+      0.17448820173740387,
+      0.18207435309886932,
+      0.17394252121448517,
+      0.17870847880840302,
+      0.1585991531610489,
+      0.17257511615753174,
+      0.18372474610805511,
+      0.1582813858985901,
+      0.1431683897972107,
+      0.14724594354629517,
+      0.15524940192699432,
+      0.12837021052837372,
+      0.15147118270397186,
+      0.13836684823036194,
+      0.14143264293670654,
+      0.1055116131901741,
+      0.12111660093069077,
+      0.1088944599032402,
+      0.12397897243499756,
+      0.12244991213083267,
+      0.11925437301397324,
+      0.11537864059209824,
+      0.12550200521945953
     ],
     "zipf_loss": [
-      4.37274169921875,
-      1.8046292066574097,
-      1.7087011337280273,
-      1.6968810558319092,
-      1.6936335563659668,
-      1.690199375152588,
-      1.6878787279129028,
-      1.6839007139205933,
-      1.6731270551681519,
-      1.5985805988311768,
-      1.4745081663131714,
-      1.278465986251831,
-      1.1647508144378662,
-      1.0889580249786377,
-      0.9403650760650635,
-      0.8638197183609009,
-      0.8192105293273926,
-      0.7208845019340515,
-      0.5981146097183228,
-      0.33281421661376953,
-      0.15102912485599518,
-      0.06269875168800354,
-      0.02885546162724495,
-      0.016147876158356667,
-      0.009403230622410774,
-      0.010388587601482868,
-      0.031447142362594604,
-      0.011647149920463562,
-      0.008088257163763046,
-      0.006815792992711067,
-      0.009629786014556885,
-      0.01600518450140953,
-      0.004471144638955593,
-      0.010364695452153683,
-      0.00525753665715456,
-      0.008417721837759018,
-      0.010751370340585709,
-      0.007105482742190361,
-      0.0021849004551768303,
-      0.005415467079728842,
-      0.006301610730588436,
-      0.00802694447338581,
-      0.012695873156189919,
-      0.0035742726176977158,
-      0.016842156648635864,
-      0.01661541312932968,
-      0.009708529338240623,
-      0.008232427760958672,
-      0.007224865257740021,
-      0.002215202199295163,
-      0.004029060248285532,
-      0.004625140223652124,
-      0.005678504705429077,
-      0.0038065649569034576,
-      0.004278676584362984,
-      0.0037967078387737274,
-      0.0046645235270261765,
-      0.0013988213613629341,
-      0.004286607727408409,
-      0.0054692612029612064,
-      0.002797577530145645,
-      0.0038844943046569824,
-      0.006156958639621735,
-      0.00890234112739563,
-      0.0011534709483385086,
-      0.0036813169717788696,
-      0.0040760282427072525,
-      0.004231432918459177,
-      0.003363661468029022,
-      0.0017783557996153831
+      8.190372467041016,
+      4.21774959564209,
+      2.0889029502868652,
+      1.7756659984588623,
+      1.715381383895874,
+      1.7015979290008545,
+      1.6966168880462646,
+      1.6921616792678833,
+      1.6919775009155273,
+      1.6897456645965576,
+      1.6901496648788452,
+      1.690834403038025,
+      1.696687936782837,
+      1.7070481777191162,
+      1.7052288055419922,
+      1.7003854513168335,
+      1.6773663759231567,
+      1.6252772808074951,
+      1.540981650352478,
+      1.5055437088012695,
+      1.4082492589950562,
+      1.3143138885498047,
+      1.2227977514266968,
+      1.1722016334533691,
+      1.0987006425857544,
+      1.0058059692382812,
+      0.9411389231681824,
+      0.9186881184577942,
+      0.8159641623497009,
+      0.7451286315917969,
+      0.6992727518081665,
+      0.6709651947021484,
+      0.6331524848937988,
+      0.5614780187606812,
+      0.6249133348464966,
+      0.5249479413032532,
+      0.4748857021331787,
+      0.4279439449310303,
+      0.4149409234523773,
+      0.37185347080230713,
+      0.3879944980144501,
+      0.4145728647708893,
+      0.33193638920783997,
+      0.40620243549346924,
+      0.3544418215751648,
+      0.3280959725379944,
+      0.3559904098510742,
+      0.32354503870010376,
+      0.2835143208503723,
+      0.24926838278770447,
+      0.2441646009683609,
+      0.2400692105293274,
+      0.21169903874397278,
+      0.1923774778842926,
+      0.25739187002182007,
+      0.1864461600780487,
+      0.1969538927078247,
+      0.20113880932331085,
+      0.16653358936309814,
+      0.208352729678154,
+      0.16804218292236328,
+      0.1555454134941101,
+      0.17147725820541382,
+      0.15252524614334106,
+      0.1460340917110443,
+      0.15528786182403564,
+      0.13820970058441162,
+      0.1327650099992752,
+      0.15823212265968323,
+      0.1449219286441803
     ],
     "denoise_loss": [],
     "ortho_loss": [
-      0.3933292031288147,
-      0.19537021219730377,
-      0.1488562375307083,
-      0.13526654243469238,
-      0.1290234625339508,
-      0.12106598168611526,
-      0.13526804745197296,
-      0.1508757472038269,
-      0.21096870303153992,
-      0.2514957785606384,
-      0.27831581234931946,
-      0.2901284098625183,
-      0.29795151948928833,
-      0.32490643858909607,
-      0.3221665322780609,
-      0.3283940255641937,
-      0.3153337240219116,
-      0.3107602894306183,
-      0.30619168281555176,
-      0.3000865578651428,
-      0.29851052165031433,
-      0.29726409912109375,
-      0.2884874641895294,
-      0.28946927189826965,
-      0.27607646584510803,
-      0.28382158279418945,
-      0.28244829177856445,
-      0.29492366313934326,
-      0.28803184628486633,
-      0.2858043313026428,
-      0.27292492985725403,
-      0.2727271318435669,
-      0.27270883321762085,
-      0.277759850025177,
-      0.28240156173706055,
-      0.2703619599342346,
-      0.2714839279651642,
-      0.27064526081085205,
-      0.2735464572906494,
-      0.2756211757659912,
-      0.28339874744415283,
-      0.27706512808799744,
-      0.2764289081096649,
-      0.2683185040950775,
-      0.2532169818878174,
-      0.24946703016757965,
-      0.2466099113225937,
-      0.2510632872581482,
-      0.2696925401687622,
-      0.2670671343803406,
-      0.2715094983577728,
-      0.26723748445510864,
-      0.26079726219177246,
-      0.2603513300418854,
-      0.25771471858024597,
-      0.25748154520988464,
-      0.2570980489253998,
-      0.2605780363082886,
-      0.26496967673301697,
-      0.2605254054069519,
-      0.25853437185287476,
-      0.2556053102016449,
-      0.24565964937210083,
-      0.2401438057422638,
-      0.23765115439891815,
-      0.23894348740577698,
-      0.23828937113285065,
-      0.23862119019031525,
-      0.23789578676223755,
-      0.23580653965473175
+      0.6088240742683411,
+      0.33010056614875793,
+      0.22861000895500183,
+      0.21283285319805145,
+      0.17951442301273346,
+      0.16018734872341156,
+      0.11169280111789703,
+      0.12691719830036163,
+      0.1456661969423294,
+      0.15746210515499115,
+      0.1443333923816681,
+      0.16156138479709625,
+      0.20633621513843536,
+      0.22460925579071045,
+      0.26161104440689087,
+      0.27444061636924744,
+      0.28116294741630554,
+      0.28717944025993347,
+      0.3055819571018219,
+      0.325307697057724,
+      0.3221662938594818,
+      0.3159293830394745,
+      0.3048442602157593,
+      0.30672410130500793,
+      0.301876962184906,
+      0.29914870858192444,
+      0.31866511702537537,
+      0.33082228899002075,
+      0.33669188618659973,
+      0.32973605394363403,
+      0.32963281869888306,
+      0.33795875310897827,
+      0.3385995030403137,
+      0.3364645540714264,
+      0.33232346177101135,
+      0.33818763494491577,
+      0.3330787122249603,
+      0.3314555287361145,
+      0.33205369114875793,
+      0.3319832682609558,
+      0.32697418332099915,
+      0.3278193771839142,
+      0.3068898022174835,
+      0.309592068195343,
+      0.3145006597042084,
+      0.3080871105194092,
+      0.3092295825481415,
+      0.3054082989692688,
+      0.30440065264701843,
+      0.299023300409317,
+      0.3067370355129242,
+      0.31260034441947937,
+      0.3118235766887665,
+      0.3144630491733551,
+      0.31215304136276245,
+      0.31217923760414124,
+      0.31190818548202515,
+      0.3100627064704895,
+      0.3103569447994232,
+      0.3121229112148285,
+      0.3065374791622162,
+      0.3073209226131439,
+      0.3084685206413269,
+      0.30692750215530396,
+      0.3080828785896301,
+      0.30695533752441406,
+      0.3065243661403656,
+      0.30688145756721497,
+      0.3082342743873596,
+      0.30828753113746643
     ],
     "lr": [
-      7.840000000000001e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      7.864766839378239e-05,
-      7.63160621761658e-05,
-      7.398445595854923e-05,
-      7.165284974093265e-05,
-      6.932124352331606e-05,
-      6.69896373056995e-05,
-      6.465803108808292e-05,
-      6.041450777202072e-05,
-      5.8082901554404154e-05,
-      5.5751295336787566e-05,
-      5.3419689119171e-05,
-      5.108808290155441e-05,
-      4.8756476683937825e-05,
-      4.642487046632125e-05,
-      4.218134715025906e-05,
-      3.98497409326425e-05,
-      3.7518134715025914e-05,
-      3.518652849740933e-05,
-      3.285492227979275e-05,
-      3.0523316062176166e-05,
-      2.8191709844559595e-05,
-      2.3948186528497416e-05,
-      2.1616580310880825e-05,
-      1.9284974093264255e-05,
-      1.6953367875647667e-05,
-      1.4621761658031097e-05,
-      1.2290155440414508e-05,
-      9.958549222797919e-06
+      1.6752136752136756e-05,
+      3.384615384615385e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      3.9947798576324814e-05,
+      3.8761402583706826e-05,
+      3.757500659108885e-05,
+      3.6388610598470864e-05,
+      3.5202214605852884e-05,
+      3.401581861323491e-05,
+      3.282942262061693e-05,
+      3.0670181914052204e-05,
+      2.948378592143422e-05,
+      2.8297389928816243e-05,
+      2.711099393619826e-05,
+      2.5924597943580284e-05,
+      2.4738201950962303e-05,
+      2.3551805958344316e-05,
+      2.1392565251779595e-05,
+      2.020616925916161e-05,
+      1.901977326654364e-05,
+      1.783337727392566e-05,
+      1.6646981281307675e-05,
+      1.546058528868969e-05,
+      1.427418929607171e-05,
+      1.2114948589506984e-05,
+      1.0928552596889013e-05,
+      9.742156604271029e-06,
+      8.555760611653046e-06,
+      7.369364619035064e-06,
+      6.182968626417082e-06,
+      4.996572633799099e-06
     ],
     "emb_lr": [],
     "eval_step": [
@@ -591,7 +591,7 @@
       3869
     ],
     "eval_accuracy": [
-      0.0,
+      0.01,
       0.0,
       0.0,
       0.0,
@@ -603,508 +603,508 @@
       0.0
     ]
   },
-  "final_accuracy": 0.9707142857142858,
+  "final_accuracy": 0.7783333333333333,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
       "K": null,
       "mode": "sft",
       "n_digits": 6,
-      "n_per_split": 50
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.98,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 295
+            "accuracy": 0.996694214876033,
+            "count": 605
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 55
+            "count": 95
           }
         }
       },
       "add_S1": {
-        "full_accuracy": 0.98,
-        "n_examples": 50,
+        "full_accuracy": 0.86,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 126
+            "accuracy": 0.9754901960784313,
+            "count": 204
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 79
+            "accuracy": 0.9763313609467456,
+            "count": 169
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 21
+            "count": 31
           },
           "UC": {
-            "accuracy": 0.9919354838709677,
-            "count": 124
+            "accuracy": 0.9831081081081081,
+            "count": 296
           }
         }
       },
       "add_S2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.84,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 75
+            "count": 163
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 62
+            "accuracy": 0.9615384615384616,
+            "count": 130
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
+            "count": 87
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 111
+            "accuracy": 0.9408866995073891,
+            "count": 203
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 63
+            "accuracy": 0.9914529914529915,
+            "count": 117
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.33,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 60
+            "accuracy": 0.9834710743801653,
+            "count": 121
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 57
+            "accuracy": 0.9834710743801653,
+            "count": 121
           },
           "SS": {
-            "accuracy": 1.0,
-            "count": 19
+            "accuracy": 0.9591836734693877,
+            "count": 49
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 104
+            "accuracy": 0.6505376344086021,
+            "count": 186
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 110
+            "accuracy": 0.9147982062780269,
+            "count": 223
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 0.88,
-        "n_examples": 50,
+        "full_accuracy": 0.29,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 48
+            "accuracy": 0.9903846153846154,
+            "count": 104
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 52
+            "accuracy": 0.9716981132075472,
+            "count": 106
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 7
+            "count": 23
           },
           "UC": {
-            "accuracy": 0.9325842696629213,
-            "count": 89
+            "accuracy": 0.61875,
+            "count": 160
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 154
+            "accuracy": 0.7100977198697068,
+            "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.92,
-        "n_examples": 50,
+        "full_accuracy": 0.14,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.96,
-            "count": 50
+            "accuracy": 0.18,
+            "count": 100
           },
           "US": {
-            "accuracy": 0.99,
-            "count": 200
+            "accuracy": 0.4975,
+            "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.28,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.42,
+            "count": 100
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 250
+            "accuracy": 0.526,
+            "count": 500
           }
         }
       },
       "add_random": {
-        "full_accuracy": 1.0,
+        "full_accuracy": 0.88,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 431
+            "accuracy": 0.9888143176733781,
+            "count": 447
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 316
+            "accuracy": 0.9875,
+            "count": 320
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
+            "count": 56
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 560
+            "accuracy": 0.9716446124763705,
+            "count": 529
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 54
+            "accuracy": 0.9791666666666666,
+            "count": 48
           }
         }
       },
       "add_C3": {
-        "full_accuracy": 0.98,
-        "n_examples": 50,
+        "full_accuracy": 0.62,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 150
+            "count": 300
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.99,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.9903846153846154,
-            "count": 104
+            "accuracy": 0.8082901554404145,
+            "count": 193
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 46
+            "accuracy": 0.9158878504672897,
+            "count": 107
           }
         }
       },
       "add_C4": {
-        "full_accuracy": 0.96,
-        "n_examples": 50,
+        "full_accuracy": 0.65,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.983739837398374,
-            "count": 123
+            "accuracy": 0.875,
+            "count": 256
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 77
+            "accuracy": 0.8819444444444444,
+            "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.98,
-        "n_examples": 50,
+        "full_accuracy": 0.52,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.95,
+            "count": 100
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.99,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.9935064935064936,
-            "count": 154
+            "accuracy": 0.8496732026143791,
+            "count": 306
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 96
+            "accuracy": 0.8350515463917526,
+            "count": 194
           }
         }
       },
       "add_C6": {
-        "full_accuracy": 0.94,
-        "n_examples": 50,
+        "full_accuracy": 0.56,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.9835164835164835,
-            "count": 182
+            "accuracy": 0.8852459016393442,
+            "count": 366
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 118
+            "accuracy": 0.905982905982906,
+            "count": 234
           }
         }
       },
       "sub_M0": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.91,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 294
+            "accuracy": 0.9850249584026622,
+            "count": 601
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 99
           }
         }
       },
       "sub_M1": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.94,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 143
+            "accuracy": 0.996415770609319,
+            "count": 279
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 69
+            "accuracy": 0.9862068965517241,
+            "count": 145
           },
           "ME": {
-            "accuracy": 1.0,
-            "count": 15
+            "accuracy": 0.9583333333333334,
+            "count": 24
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 123
+            "accuracy": 0.9920634920634921,
+            "count": 252
           }
         }
       },
       "sub_M2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.85,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 108
+            "accuracy": 0.9906103286384976,
+            "count": 213
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 52
+            "accuracy": 0.9823008849557522,
+            "count": 113
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 85
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 87
+            "accuracy": 0.9281767955801105,
+            "count": 181
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 51
+            "count": 108
           }
         }
       },
       "sub_M3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.3,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 94
+            "count": 179
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 51
+            "accuracy": 0.9805825242718447,
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 25
+            "count": 56
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 78
+            "accuracy": 0.6040268456375839,
+            "count": 149
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 102
+            "accuracy": 0.9061032863849765,
+            "count": 213
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 0.48,
-        "n_examples": 50,
+        "full_accuracy": 0.03,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.48,
-            "count": 50
+            "accuracy": 0.43,
+            "count": 100
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 150
+            "accuracy": 0.49333333333333335,
+            "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.6,
-        "n_examples": 50,
+        "full_accuracy": 0.06,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.94,
-            "count": 50
+            "accuracy": 0.41,
+            "count": 100
           },
           "UD": {
-            "accuracy": 0.9,
-            "count": 200
+            "accuracy": 0.4075,
+            "count": 400
           }
         }
       },
       "sub_random": {
-        "full_accuracy": 1.0,
+        "full_accuracy": 0.875,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 588
+            "accuracy": 0.99,
+            "count": 600
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 268
+            "accuracy": 0.9887640449438202,
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 60
+            "count": 53
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 447
+            "accuracy": 0.9635535307517085,
+            "count": 439
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 37
+            "accuracy": 0.975609756097561,
+            "count": 41
           }
         }
       },
       "sub_B3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.75,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 150
+            "count": 300
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 107
+            "accuracy": 0.883248730964467,
+            "count": 197
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 43
+            "accuracy": 0.9514563106796117,
+            "count": 103
           }
         }
       },
       "sub_B4": {
-        "full_accuracy": 0.88,
-        "n_examples": 50,
+        "full_accuracy": 0.55,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.9473684210526315,
-            "count": 114
+            "accuracy": 0.8421052631578947,
+            "count": 247
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 86
+            "accuracy": 0.869281045751634,
+            "count": 153
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.84,
-        "n_examples": 50,
+        "full_accuracy": 0.48,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.9738562091503268,
-            "count": 153
+            "accuracy": 0.8624161073825504,
+            "count": 298
           },
           "UD": {
-            "accuracy": 0.9484536082474226,
-            "count": 97
+            "accuracy": 0.8118811881188119,
+            "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.9442857142857143,
-      "total_examples": 1400,
+      "overall_accuracy": 0.6016666666666667,
+      "total_examples": 2400,
       "n_splits": 22
     }
   },
@@ -1114,504 +1114,504 @@
       "K": 1,
       "mode": "sorl",
       "n_digits": 6,
-      "n_per_split": 50
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 295
+            "count": 605
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 55
+            "count": 95
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 126
+            "count": 204
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 79
+            "count": 169
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 21
+            "count": 31
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 124
+            "count": 296
           }
         }
       },
       "add_S2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.96,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 75
+            "accuracy": 0.9938650306748467,
+            "count": 163
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 62
+            "accuracy": 0.9923076923076923,
+            "count": 130
           },
           "SS": {
-            "accuracy": 1.0,
-            "count": 39
+            "accuracy": 0.9770114942528736,
+            "count": 87
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 111
+            "accuracy": 0.9901477832512315,
+            "count": 203
           },
           "US": {
             "accuracy": 1.0,
-            "count": 63
+            "count": 117
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.67,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 60
+            "count": 121
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 57
+            "accuracy": 0.9917355371900827,
+            "count": 121
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 19
+            "count": 49
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 104
+            "accuracy": 0.8279569892473119,
+            "count": 186
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 110
+            "accuracy": 0.9955156950672646,
+            "count": 223
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.45,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 48
+            "count": 104
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 106
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 7
+            "count": 23
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 89
+            "accuracy": 0.6875,
+            "count": 160
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 154
+            "accuracy": 0.8273615635179153,
+            "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.62,
-        "n_examples": 50,
+        "full_accuracy": 0.21,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.62,
-            "count": 50
+            "accuracy": 0.33,
+            "count": 100
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 200
+            "accuracy": 0.6175,
+            "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.19,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.34,
+            "count": 100
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 250
+            "accuracy": 0.54,
+            "count": 500
           }
         }
       },
       "add_random": {
-        "full_accuracy": 1.0,
+        "full_accuracy": 0.97,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 431
+            "accuracy": 0.9888143176733781,
+            "count": 447
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 316
+            "count": 320
           },
           "SS": {
-            "accuracy": 1.0,
-            "count": 39
+            "accuracy": 0.9642857142857143,
+            "count": 56
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 560
+            "accuracy": 0.998109640831758,
+            "count": 529
           },
           "US": {
             "accuracy": 1.0,
-            "count": 54
+            "count": 48
           }
         }
       },
       "add_C3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.82,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 150
+            "count": 300
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 104
+            "accuracy": 0.9067357512953368,
+            "count": 193
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 46
+            "accuracy": 0.9906542056074766,
+            "count": 107
           }
         }
       },
       "add_C4": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.79,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 123
+            "accuracy": 0.92578125,
+            "count": 256
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 77
+            "accuracy": 0.9583333333333334,
+            "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.78,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 154
+            "accuracy": 0.9379084967320261,
+            "count": 306
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 96
+            "accuracy": 0.865979381443299,
+            "count": 194
           }
         }
       },
       "add_C6": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.85,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 182
+            "accuracy": 0.9617486338797814,
+            "count": 366
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 118
+            "accuracy": 0.9658119658119658,
+            "count": 234
           }
         }
       },
       "sub_M0": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.96,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 294
+            "accuracy": 0.9933444259567388,
+            "count": 601
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 99
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 143
+            "count": 279
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 69
+            "count": 145
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 15
+            "count": 24
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 123
+            "count": 252
           }
         }
       },
       "sub_M2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.97,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 108
+            "accuracy": 0.9953051643192489,
+            "count": 213
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 113
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 85
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 87
+            "accuracy": 0.988950276243094,
+            "count": 181
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 51
+            "count": 108
           }
         }
       },
       "sub_M3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.73,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 94
+            "count": 179
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 51
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 25
+            "count": 56
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 78
+            "accuracy": 0.8322147651006712,
+            "count": 149
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 102
+            "accuracy": 0.9906103286384976,
+            "count": 213
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.48,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 50
+            "accuracy": 0.62,
+            "count": 100
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 150
+            "accuracy": 0.8666666666666667,
+            "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.58,
-        "n_examples": 50,
+        "full_accuracy": 0.35,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.58,
-            "count": 50
+            "accuracy": 0.68,
+            "count": 100
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 200
+            "accuracy": 0.755,
+            "count": 400
           }
         }
       },
       "sub_random": {
-        "full_accuracy": 1.0,
+        "full_accuracy": 0.985,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 588
+            "accuracy": 0.9966666666666667,
+            "count": 600
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 268
+            "accuracy": 0.9962546816479401,
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 60
+            "count": 53
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 447
+            "count": 439
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 37
+            "count": 41
           }
         }
       },
       "sub_B3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.9,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 150
+            "accuracy": 0.9966666666666667,
+            "count": 300
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 107
+            "accuracy": 0.9543147208121827,
+            "count": 197
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 43
+            "accuracy": 0.9902912621359223,
+            "count": 103
           }
         }
       },
       "sub_B4": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.87,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 114
+            "accuracy": 0.9554655870445344,
+            "count": 247
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 86
+            "accuracy": 0.954248366013072,
+            "count": 153
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.79,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 153
+            "accuracy": 0.9328859060402684,
+            "count": 298
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 97
+            "accuracy": 0.9554455445544554,
+            "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.9707142857142858,
-      "total_examples": 1400,
+      "overall_accuracy": 0.7783333333333333,
+      "total_examples": 2400,
       "n_splits": 22
     }
   },
-  "sorl_overall_accuracy": 0.9707142857142858,
-  "sft_overall_accuracy": 0.9442857142857143
+  "sorl_overall_accuracy": 0.7783333333333333,
+  "sft_overall_accuracy": 0.6016666666666667
 }

add_sub_sorl_v1_abs10_K1_25K/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ccab70b2517264fbfc0cc6c61a4eb7e47555f03c4db8db8c06ccd93ca673eac1
+oid sha256:a3abff1d3d65f28a467e0107b3021da4659af9733b2521b6687ffa6fbb31e60f
 size 650303660

add_sub_sorl_v1_abs10_K1_25K/train_config.json

@@ -1,35 +1,84 @@
 {
-  "mode": "sorl",
-  "ops": "add_sub",
-  "n_digits": 6,
-  "n_layer": 2,
-  "n_head": 3,
-  "n_embd": 510,
-  "abs_vocab": 10,
+  "num_rollouts": 4,
   "K": 1,
+  "max_iterations": 2,
+  "memory_span_abs": 1792,
+  "memory_span_traj": 1792,
+  "temperature": 1.0,
+  "ar_search": false,
+  "response_only_abs": false,
   "alpha_info_gain": 10.0,
   "alpha_abs": 0.1,
   "alpha_soft_zipf": 1.0,
+  "alpha_ortho": 0.0,
+  "alpha_anchor": 0.0,
+  "alpha_jacobi": 0.0,
+  "decay": 0.8,
+  "target_vocab_util": 0.8,
+  "min_abs_ppl": 0.0,
+  "zipf_alpha": 1.0,
+  "lr": 4e-05,
+  "emb_lr_mult": 1.0,
+  "weight_decay": 0.01,
+  "warmup_steps": 117,
+  "cooldown_frac": 0.4,
+  "max_grad_norm": 1.0,
+  "vq_abs_pretrain_steps": 0,
+  "vq_abs_pretrain_lr": 0.001,
+  "vq_abs_pretrain_layer": -1,
+  "vq_abs_pretrain_batch_size": 256,
+  "vq_abs_pretrain_target_vectors": 20000,
   "batch_size": 64,
+  "gradient_accumulation_steps": 1,
   "num_epochs": 10,
-  "dataset_size": 25000,
-  "lr": 8e-05,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 390,
+  "save_every": 999999,
+  "eval_samples": 100,
   "output_dir": "ckpt/sweep/as_sorl_abs10_K1_25K",
+  "eval_K": 4,
+  "alpha_traj": 0.0,
+  "corrupt_method": "shuffle",
+  "corrupt_ratio": 0.3,
+  "alpha_contrastive": 1.0,
+  "gamma_contrastive": 0.5,
+  "alpha_masked_traj": 0.0,
+  "mask_nl_ratio": 0.3,
+  "mask_nl_mode": "fixed",
+  "mask_nl_fixed_id": 0,
+  "use_ste": true,
+  "n_inner": 1,
+  "random_K": null,
+  "strip_suffix": null,
+  "compress_prefix": null,
+  "random_mem_span": null,
+  "warmup_ratio": 0.03,
+  "beta2": 0.999,
+  "seed": 42,
+  "n_digits": 6,
+  "n_layer": 2,
+  "n_head": 3,
+  "n_embd": 510,
+  "ops": "add_sub",
+  "abs_vocab": 10,
+  "dataset_size": 25000,
+  "mode": "sorl",
   "device": "cuda",
   "push_to_hub": true,
   "no_wandb": false,
   "n_params": 162499262,
   "run_name": "add_sub_sorl_v1_abs10_K1_25K",
-  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
-  "timestamp": "2026-04-12T01:58:11.626522+00:00",
+  "git_commit": "78d46f8665a87f4b44bd5894bd34f393f2dea51f",
+  "timestamp": "2026-04-12T08:59:10.014670+00:00",
   "tokenizer": "Qwen/Qwen3-0.6B",
   "dataset_repo": "thoughtworks/arithmetic-sorl-data",
   "dataset_config": "add_sub_6digit",
   "model_repo": "thoughtworks/arithmetic-sorl",
   "trainer_version": "v1",
-  "wandb_run_id": "nv7szjmr",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/nv7szjmr",
-  "final_accuracy": 0.9707142857142858,
-  "sft_accuracy": 0.9442857142857143,
+  "wandb_run_id": "je0pfgat",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/je0pfgat",
+  "final_accuracy": 0.7783333333333333,
+  "sft_accuracy": 0.6016666666666667,
   "eval_method": "ArithmeticEvaluator"
 }