Upload add_sub_sorl_v1_abs50_50K

Browse files

Files changed (3) hide show

add_sub_sorl_v1_abs50_50K/metrics.json +1378 -1378
add_sub_sorl_v1_abs50_50K/model.safetensors +1 -1
add_sub_sorl_v1_abs50_50K/train_config.json +64 -15

add_sub_sorl_v1_abs50_50K/metrics.json CHANGED Viewed

@@ -153,1069 +153,1069 @@
       7788
     ],
     "loss": [
-      7.256625652313232,
-      3.0191478729248047,
-      2.4123916625976562,
-      2.1674225330352783,
-      2.1556787490844727,
-      1.8286030292510986,
-      2.14564847946167,
-      1.900666356086731,
-      1.652380108833313,
-      1.6646597385406494,
-      1.5041191577911377,
-      1.546173334121704,
-      0.7197972536087036,
-      0.08199936151504517,
-      -3.1339755058288574,
-      -10.444205284118652,
-      -10.103414535522461,
-      -12.214149475097656,
-      -12.902936935424805,
-      -13.465953826904297,
-      -13.666388511657715,
-      -14.849788665771484,
-      -14.489677429199219,
-      -14.085216522216797,
-      -15.04929256439209,
-      -14.683259010314941,
-      -13.91379451751709,
-      -15.094310760498047,
-      -13.880881309509277,
-      -14.374068260192871,
-      -14.42313003540039,
-      -14.229098320007324,
-      -15.09312629699707,
-      -14.730694770812988,
-      -14.6654634475708,
-      -13.932817459106445,
-      -14.541162490844727,
-      -13.534533500671387,
-      -14.629840850830078,
-      -14.549479484558105,
-      -14.572128295898438,
-      -14.297953605651855,
-      -14.748451232910156,
-      -14.658515930175781,
-      -14.846949577331543,
-      -14.570510864257812,
-      -14.642899513244629,
-      -15.091267585754395,
-      -14.61697769165039,
-      -14.35709285736084,
-      -14.233384132385254,
-      -13.99492073059082,
-      -13.887598037719727,
-      -13.822426795959473,
-      -13.868623733520508,
-      -13.392026901245117,
-      -14.518260955810547,
-      -12.745328903198242,
-      -9.385150909423828,
-      -8.4232177734375,
-      -5.4400715827941895,
-      -5.325277805328369,
-      -5.132301330566406,
-      -4.313063621520996,
-      -3.695901870727539,
-      -3.7530672550201416,
-      -3.7038750648498535,
-      -3.661875009536743,
-      -3.2753937244415283,
-      -3.360624074935913,
-      -2.504333257675171,
-      -3.0331249237060547,
-      -3.0124270915985107,
-      -2.910573720932007,
-      -3.7368786334991455,
-      -2.7395472526550293,
-      -2.6733858585357666,
-      -2.886018991470337,
-      -2.6889913082122803,
-      -2.5152089595794678,
-      -2.2789855003356934,
-      -2.7361295223236084,
-      -2.4921624660491943,
-      -2.7641701698303223,
-      -4.348557949066162,
-      -2.480908155441284,
-      -2.207150936126709,
-      -2.63488507270813,
-      -2.250746726989746,
-      -2.611670970916748,
-      -2.541576385498047,
-      -2.34275484085083,
-      -2.689805746078491,
-      -2.281898021697998,
-      -2.7649309635162354,
-      -2.1240293979644775,
-      -2.607470750808716,
-      -2.4728121757507324,
-      -2.0544490814208984,
-      -2.0300307273864746,
-      -2.1023619174957275,
-      -1.887721061706543,
-      -2.0482027530670166,
-      -1.9892445802688599,
-      -1.9722803831100464,
-      -2.084959030151367,
-      -2.446387529373169,
-      -1.835938811302185,
-      -2.5824408531188965,
-      -2.1360812187194824,
-      -1.9915828704833984,
-      -1.6431983709335327,
-      -1.4731082916259766,
-      -1.7144525051116943,
-      -1.7019168138504028,
-      -1.6519891023635864,
-      -1.7413341999053955,
-      -1.588895320892334,
-      -1.539756417274475,
-      -1.904645562171936,
-      -1.5546034574508667,
-      -1.7151820659637451,
-      -1.450798511505127,
-      -1.574430227279663,
-      -1.8458795547485352,
-      -1.5194255113601685,
-      -1.6603127717971802,
-      -2.0166845321655273,
-      -1.2642289400100708,
-      -1.7409981489181519,
-      -1.6157901287078857,
-      -1.5526642799377441,
-      -1.2320806980133057,
-      -1.5149484872817993,
-      -1.4476909637451172,
-      -1.6797325611114502,
-      -1.3895442485809326,
-      -1.231076717376709,
-      -1.2591882944107056,
-      -1.7371399402618408,
-      -1.5310636758804321,
-      -1.4542627334594727,
-      -1.2413599491119385,
-      -1.226067304611206,
-      -1.6244958639144897,
-      -1.3765522241592407,
-      -1.362652063369751,
-      -1.392403244972229,
-      -1.3551605939865112,
-      -1.0807950496673584
     ],
     "base_loss": [
-      6.029807090759277,
-      2.3067009449005127,
-      1.9342564344406128,
-      1.8905317783355713,
-      1.8440930843353271,
-      1.8317269086837769,
-      1.8801648616790771,
-      1.8195487260818481,
-      1.7616671323776245,
-      1.7560369968414307,
-      1.7867367267608643,
-      1.7833213806152344,
-      1.7948106527328491,
-      1.7519046068191528,
-      1.948911428451538,
-      2.001823902130127,
-      1.8075017929077148,
-      1.8323737382888794,
-      1.797086477279663,
-      1.8139381408691406,
-      1.751043438911438,
-      1.8313592672348022,
-      1.751556634902954,
-      1.7140840291976929,
-      1.789836049079895,
-      1.7125715017318726,
-      1.6791929006576538,
-      1.7678550481796265,
-      1.6537859439849854,
-      1.6602624654769897,
-      1.6753387451171875,
-      1.6629526615142822,
-      1.7369420528411865,
-      1.683754563331604,
-      1.6948719024658203,
-      1.6238089799880981,
-      1.6669832468032837,
-      1.5685263872146606,
-      1.6912544965744019,
-      1.6462609767913818,
-      1.6818253993988037,
-      1.636511206626892,
-      1.6857174634933472,
-      1.669575810432434,
-      1.6751861572265625,
-      1.6484739780426025,
-      1.6467863321304321,
-      1.6952067613601685,
-      1.6533762216567993,
-      1.616503357887268,
-      1.6140310764312744,
-      1.5757274627685547,
-      1.5750893354415894,
-      1.556767225265503,
-      1.5602978467941284,
-      1.515640139579773,
-      1.6252305507659912,
-      1.4415589570999146,
-      1.0609798431396484,
-      0.9612196683883667,
-      0.616158127784729,
-      0.6064944863319397,
-      0.5862141251564026,
-      0.4924693703651428,
-      0.4313984513282776,
-      0.42700475454330444,
-      0.42180582880973816,
-      0.41957685351371765,
-      0.37408900260925293,
-      0.38498300313949585,
-      0.28954702615737915,
-      0.348208487033844,
-      0.3462330996990204,
-      0.3341756761074066,
-      0.4224683344364166,
-      0.3134305477142334,
-      0.3083657920360565,
-      0.3302149474620819,
-      0.30693626403808594,
-      0.2877400815486908,
-      0.2639155685901642,
-      0.3185356557369232,
-      0.2856982350349426,
-      0.31456050276756287,
-      0.49335941672325134,
-      0.28367578983306885,
-      0.25294229388237,
-      0.30259618163108826,
-      0.2595691978931427,
-      0.30289167165756226,
-      0.291527658700943,
-      0.26810169219970703,
-      0.3059633672237396,
-      0.2611132562160492,
-      0.3158523738384247,
-      0.2444075345993042,
-      0.29604029655456543,
-      0.2825319468975067,
-      0.23427459597587585,
-      0.23259370028972626,
-      0.2404058873653412,
-      0.21756012737751007,
-      0.23280969262123108,
-      0.22855813801288605,
-      0.2257063388824463,
-      0.23912663757801056,
-      0.27966248989105225,
-      0.21003004908561707,
-      0.29382458329200745,
-      0.2435278445482254,
-      0.2295869141817093,
-      0.18904468417167664,
-      0.17142455279827118,
-      0.1948322355747223,
-      0.19676467776298523,
-      0.19100677967071533,
-      0.19870205223560333,
-      0.18229179084300995,
-      0.17854009568691254,
-      0.2182820737361908,
-      0.1792173683643341,
-      0.19656817615032196,
-      0.16844725608825684,
-      0.18259193003177643,
-      0.21280623972415924,
-      0.17412368953227997,
-      0.19087910652160645,
-      0.22902394831180573,
-      0.14896346628665924,
-      0.20004956424236298,
-      0.18591725826263428,
-      0.17941395938396454,
-      0.14243106544017792,
-      0.17380957305431366,
-      0.16619420051574707,
-      0.19304493069648743,
-      0.15973064303398132,
-      0.1419793665409088,
-      0.1448678970336914,
-      0.2004563808441162,
-      0.17605385184288025,
-      0.16834628582000732,
-      0.14428769052028656,
-      0.14093129336833954,
-      0.18535971641540527,
-      0.1585533171892166,
-      0.1568572074174881,
-      0.15938758850097656,
-      0.15521882474422455,
-      0.126205712556839
     ],
     "info_loss": [
-      -0.20561504364013672,
-      -0.03972625732421875,
-      -0.048748135566711426,
-      -0.06750452518463135,
-      -0.06405246257781982,
-      -0.09477114677429199,
-      -0.06761360168457031,
-      -0.08672130107879639,
-      -0.10533666610717773,
-      -0.10386943817138672,
-      -0.12152564525604248,
-      -0.11574661731719971,
-      -0.19611811637878418,
-      -0.24177753925323486,
-      -0.5657413005828857,
-      -1.2865196466445923,
-      -1.2276489734649658,
-      -1.4379183053970337,
-      -1.5007377862930298,
-      -1.5571799278259277,
-      -1.5694807767868042,
-      -1.6956429481506348,
-      -1.6478774547576904,
-      -1.6037368774414062,
-      -1.7057857513427734,
-      -1.6578631401062012,
-      -1.5784200429916382,
-      -1.7046573162078857,
-      -1.569814682006836,
-      -1.6190284490585327,
-      -1.6252418756484985,
-      -1.6039395332336426,
-      -1.6967638731002808,
-      -1.6540355682373047,
-      -1.6475841999053955,
-      -1.5689774751663208,
-      -1.635087251663208,
-      -1.5238193273544312,
-      -1.6450031995773315,
-      -1.630631685256958,
-      -1.637630820274353,
-      -1.6060518026351929,
-      -1.6621246337890625,
-      -1.6464064121246338,
-      -1.6617577075958252,
-      -1.6344650983810425,
-      -1.6399197578430176,
-      -1.6889852285385132,
-      -1.639074444770813,
-      -1.6081275939941406,
-      -1.5951539278030396,
-      -1.566024661064148,
-      -1.5557743310928345,
-      -1.5495342016220093,
-      -1.5556063652038574,
-      -1.5009874105453491,
-      -1.6231387853622437,
-      -1.4322359561920166,
-      -1.0561802387237549,
-      -0.9466190338134766,
-      -0.6137336492538452,
-      -0.6038743257522583,
-      -0.5819218754768372,
-      -0.4909442663192749,
-      -0.4217419922351837,
-      -0.4254307448863983,
-      -0.4203161597251892,
-      -0.4181394577026367,
-      -0.3734082281589508,
-      -0.38232800364494324,
-      -0.2871572971343994,
-      -0.34667515754699707,
-      -0.3433464467525482,
-      -0.3323284983634949,
-      -0.4220069944858551,
-      -0.3129675090312958,
-      -0.30473649501800537,
-      -0.32990047335624695,
-      -0.3063972592353821,
-      -0.28762325644493103,
-      -0.2610863447189331,
-      -0.312552273273468,
-      -0.2848331332206726,
-      -0.3143031597137451,
-      -0.49174460768699646,
-      -0.2831017076969147,
-      -0.25272035598754883,
-      -0.3016011118888855,
-      -0.2592092454433441,
-      -0.2988108992576599,
-      -0.2913071811199188,
-      -0.2679586410522461,
-      -0.3054676353931427,
-      -0.26094746589660645,
-      -0.3152065873146057,
-      -0.24421927332878113,
-      -0.295553594827652,
-      -0.28226152062416077,
-      -0.2340558022260666,
-      -0.23213818669319153,
-      -0.24020898342132568,
-      -0.21750272810459137,
-      -0.23276019096374512,
-      -0.22842691838741302,
-      -0.22563359141349792,
-      -0.23904889822006226,
-      -0.2794150114059448,
-      -0.20988059043884277,
-      -0.29378125071525574,
-      -0.24314871430397034,
-      -0.2293650060892105,
-      -0.18900363147258759,
-      -0.17137543857097626,
-      -0.1948050558567047,
-      -0.19659706950187683,
-      -0.19093747437000275,
-      -0.19865652918815613,
-      -0.1822470873594284,
-      -0.1785041093826294,
-      -0.21816547214984894,
-      -0.17919395864009857,
-      -0.19648776948451996,
-      -0.16839498281478882,
-      -0.18214161694049835,
-      -0.212591290473938,
-      -0.17409944534301758,
-      -0.19081123173236847,
-      -0.2289951741695404,
-      -0.14660219848155975,
-      -0.20001469552516937,
-      -0.18589811027050018,
-      -0.1793726533651352,
-      -0.14236652851104736,
-      -0.1737845093011856,
-      -0.16613999009132385,
-      -0.19301597774028778,
-      -0.15971405804157257,
-      -0.14195962250232697,
-      -0.14485248923301697,
-      -0.20041891932487488,
-      -0.17602841556072235,
-      -0.16832193732261658,
-      -0.14427153766155243,
-      -0.14090751111507416,
-      -0.18532036244869232,
-      -0.1585266888141632,
-      -0.15683983266353607,
-      -0.15936271846294403,
-      -0.1551767736673355,
-      -0.12618587911128998
     ],
     "abs_loss": [
-      3.7497520446777344,
-      2.6481971740722656,
-      2.8985848426818848,
-      2.818944215774536,
-      2.870543956756592,
-      2.872476100921631,
-      2.758188247680664,
-      2.8428196907043457,
-      2.7872250080108643,
-      2.910464286804199,
-      2.752615213394165,
-      2.7820749282836914,
-      2.7789087295532227,
-      2.835494041442871,
-      2.5386619567871094,
-      2.0468544960021973,
-      1.7976104021072388,
-      1.6245003938674927,
-      1.4929277896881104,
-      1.4750802516937256,
-      1.5771673917770386,
-      1.666552186012268,
-      1.0843603610992432,
-      1.545884132385254,
-      1.4265331029891968,
-      0.999521017074585,
-      1.2854008674621582,
-      1.195152759552002,
-      1.1227796077728271,
-      0.9406141042709351,
-      1.0983028411865234,
-      0.85012286901474,
-      0.8116261959075928,
-      0.7224507927894592,
-      0.8427834510803223,
-      0.8805156946182251,
-      0.8626446723937988,
-      0.9968369007110596,
-      0.9508752822875977,
-      0.8513461351394653,
-      0.6574622392654419,
-      0.8986561298370361,
-      0.5942917466163635,
-      0.7797585129737854,
-      0.7031676173210144,
-      0.6163696050643921,
-      0.6315547227859497,
-      0.7303791046142578,
-      0.5330150723457336,
-      0.5637519359588623,
-      0.5594895482063293,
-      0.5726730823516846,
-      0.5400386452674866,
-      0.5256966352462769,
-      0.44924086332321167,
-      0.47774025797843933,
-      0.5889784097671509,
-      0.4862280488014221,
-      0.5457887649536133,
-      0.4811607599258423,
-      0.5763522982597351,
-      0.4618305563926697,
-      0.49888068437576294,
-      0.5134003758430481,
-      0.4723852872848511,
-      0.4225727915763855,
-      0.3785339295864105,
-      0.3270162343978882,
-      0.3133487105369568,
-      0.33950677514076233,
-      0.32967162132263184,
-      0.3334130644798279,
-      0.31977397203445435,
-      0.39836356043815613,
-      0.3870335817337036,
-      0.27905893325805664,
-      0.38264474272727966,
-      0.3101273775100708,
-      0.2242627888917923,
-      0.24193796515464783,
-      0.222909078001976,
-      0.20927608013153076,
-      0.23676183819770813,
-      0.30053985118865967,
-      0.34205877780914307,
-      0.22115281224250793,
-      0.252194344997406,
-      0.2795361876487732,
-      0.2963107228279114,
-      0.26521268486976624,
-      0.3244202733039856,
-      0.28083187341690063,
-      0.18782655894756317,
-      0.273157000541687,
-      0.21937644481658936,
-      0.2191070020198822,
-      0.2309451699256897,
-      0.2086794078350067,
-      0.20153307914733887,
-      0.2183530181646347,
-      0.17671014368534088,
-      0.1841510534286499,
-      0.22086754441261292,
-      0.23169928789138794,
-      0.27007248997688293,
-      0.20892009139060974,
-      0.17916454374790192,
-      0.20730438828468323,
-      0.20521558821201324,
-      0.17557865381240845,
-      0.18920457363128662,
-      0.20690660178661346,
-      0.1799832135438919,
-      0.15656515955924988,
-      0.18862086534500122,
-      0.17772457003593445,
-      0.17366942763328552,
-      0.16974030435085297,
-      0.16747745871543884,
-      0.16128060221672058,
-      0.19063544273376465,
-      0.1916358768939972,
-      0.20035310089588165,
-      0.16092176735401154,
-      0.1909703016281128,
-      0.1657269448041916,
-      0.1640641689300537,
-      0.14971278607845306,
-      0.10842731595039368,
-      0.2196534425020218,
-      0.16537384688854218,
-      0.1963886320590973,
-      0.16384944319725037,
-      0.13870114088058472,
-      0.147671639919281,
-      0.17166978120803833,
-      0.15626773238182068,
-      0.17788034677505493,
-      0.1344836950302124,
-      0.1465415060520172,
-      0.19216865301132202,
-      0.13693669438362122,
-      0.1630391776561737,
-      0.11989884078502655,
-      0.1325468271970749,
-      0.1899355947971344,
-      0.12264394760131836,
-      0.12508927285671234,
-      0.1484333872795105,
-      0.13628548383712769
     ],
     "zipf_loss": [
-      2.907993793487549,
-      0.8448899388313293,
-      0.6757582426071167,
-      0.6700415015220642,
-      0.6650558114051819,
-      0.6573399305343628,
-      0.6658006906509399,
-      0.6640486717224121,
-      0.6653571128845215,
-      0.6562708020210266,
-      0.6573773622512817,
-      0.6421106457710266,
-      0.6082768440246582,
-      0.46432074904441833,
-      0.3206597566604614,
-      0.21448186039924622,
-      0.1858120709657669,
-      0.17020957171916962,
-      0.1580614596605301,
-      0.14439967274665833,
-      0.11965931951999664,
-      0.1086258664727211,
-      0.12910524010658264,
-      0.08348017930984497,
-      0.07607515156269073,
-      0.0828506276011467,
-      0.06267359852790833,
-      0.06489258259534836,
-      0.051201872527599335,
-      0.06189261004328728,
-      0.04411982744932175,
-      0.06233106553554535,
-      0.05640706419944763,
-      0.05366092175245285,
-      0.031228451058268547,
-      0.04509659856557846,
-      0.05646108090877533,
-      0.0354497954249382,
-      0.033848002552986145,
-      0.025441396981477737,
-      0.05660788342356682,
-      0.0361880287528038,
-      0.12764789164066315,
-      0.057995907962322235,
-      0.025125017389655113,
-      0.06402861326932907,
-      0.046355463564395905,
-      0.03033997118473053,
-      0.06708943098783493,
-      0.05130421370267868,
-      0.04817532002925873,
-      0.03233150392770767,
-      0.041051946580410004,
-      0.06357830762863159,
-      0.08221856504678726,
-      0.05443323031067848,
-      0.028998034074902534,
-      0.08684802800416946,
-      0.0610925555229187,
-      0.0336371473968029,
-      0.023472048342227936,
-      0.06078788638114929,
-      0.05081494152545929,
-      0.05256975442171097,
-      0.042881276458501816,
-      0.031977828592061996,
-      0.03962726891040802,
-      0.06724098324775696,
-      0.05326463282108307,
-      0.04372230917215347,
-      0.04472571983933449,
-      0.052076928317546844,
-      0.042826779186725616,
-      0.03869928792119026,
-      0.02201938070356846,
-      0.04879137873649597,
-      0.02734885923564434,
-      0.05175810307264328,
-      0.04561843350529671,
-      0.04908981919288635,
-      0.045671574771404266,
-      0.04992988705635071,
-      0.04679465293884277,
-      0.0342470221221447,
-      0.04132283851504326,
-      0.04431803524494171,
-      0.04189082980155945,
-      0.0505763441324234,
-      0.05214542895555496,
-      0.047025106847286224,
-      0.04752563685178757,
-      0.04064689949154854,
-      0.04012472555041313,
-      0.03914778679609299,
-      0.04934476315975189,
-      0.051845304667949677,
-      0.028930390253663063,
-      0.04640305042266846,
-      0.03168107569217682,
-      0.03692198544740677,
-      0.041651077568531036,
-      0.05133098363876343,
-      0.024502430111169815,
-      0.04329643025994301,
-      0.031341951340436935,
-      0.04551147669553757,
-      0.0501837432384491,
-      0.03210654854774475,
-      0.041025444865226746,
-      0.03432022035121918,
-      0.0535600446164608,
-      0.03710262104868889,
-      0.051223304122686386,
-      0.02310934104025364,
-      0.04842713102698326,
-      0.04860638454556465,
-      0.029162120074033737,
-      0.03430991619825363,
-      0.04999687895178795,
-      0.042598988860845566,
-      0.03905525058507919,
-      0.03396385908126831,
-      0.04466875270009041,
-      0.04830180108547211,
-      0.048130013048648834,
-      0.03087255358695984,
-      0.040513940155506134,
-      0.029271963983774185,
-      0.04198697581887245,
-      0.037133827805519104,
-      0.04073633253574371,
-      0.04200948029756546,
-      0.032768625766038895,
-      0.03521694615483284,
-      0.03274751082062721,
-      0.040215421468019485,
-      0.03223889321088791,
-      0.02875204011797905,
-      0.031020455062389374,
-      0.05193880572915077,
-      0.03394973650574684,
-      0.046916697174310684,
-      0.04076382517814636,
-      0.030086800456047058,
-      0.03009340539574623,
-      0.031167635694146156,
-      0.03662469983100891,
-      0.029327519237995148,
-      0.02654491364955902,
-      0.04122946411371231
     ],
     "denoise_loss": [],
     "ortho_loss": [
-      0.10270898044109344,
-      0.045592378824949265,
-      0.034897129982709885,
-      0.03289920464158058,
-      0.03318662568926811,
-      0.035863105207681656,
-      0.03718609735369682,
-      0.04016987606883049,
-      0.04869673773646355,
-      0.053654879331588745,
-      0.061492759734392166,
-      0.06294921040534973,
-      0.06910595297813416,
-      0.07273681461811066,
-      0.076898954808712,
-      0.0958031639456749,
-      0.1058020293712616,
-      0.11720982193946838,
-      0.13132666051387787,
-      0.1452067792415619,
-      0.15491452813148499,
-      0.16184890270233154,
-      0.17446120083332062,
-      0.18189102411270142,
-      0.1888672560453415,
-      0.19659653306007385,
-      0.1996685415506363,
-      0.20556356012821198,
-      0.2079426646232605,
-      0.21441443264484406,
-      0.2170758992433548,
-      0.22126887738704681,
-      0.22582681477069855,
-      0.22605378925800323,
-      0.2280348688364029,
-      0.2285757064819336,
-      0.22907176613807678,
-      0.23065848648548126,
-      0.23264732956886292,
-      0.234690859913826,
-      0.23757733404636383,
-      0.23970840871334076,
-      0.2384733259677887,
-      0.23657508194446564,
-      0.23852205276489258,
-      0.24019040167331696,
-      0.23936468362808228,
-      0.23304446041584015,
-      0.23387999832630157,
-      0.2349344789981842,
-      0.23946575820446014,
-      0.24182644486427307,
-      0.24514825642108917,
-      0.2487141191959381,
-      0.2516993284225464,
-      0.2516952455043793,
-      0.2530873715877533,
-      0.2569259703159332,
-      0.2646262049674988,
-      0.2720961272716522,
-      0.27863892912864685,
-      0.2827668786048889,
-      0.28322896361351013,
-      0.29026034474372864,
-      0.296204149723053,
-      0.2996149957180023,
-      0.3005698323249817,
-      0.30707406997680664,
-      0.3101491630077362,
-      0.3167986273765564,
-      0.3269317150115967,
-      0.32879483699798584,
-      0.33054113388061523,
-      0.33136507868766785,
-      0.3334370255470276,
-      0.3420204818248749,
-      0.34395188093185425,
-      0.3462192118167877,
-      0.350506067276001,
-      0.3514798581600189,
-      0.35246649384498596,
-      0.3512493073940277,
-      0.3530470132827759,
-      0.355428546667099,
-      0.359707772731781,
-      0.3637158274650574,
-      0.3649544417858124,
-      0.3662114441394806,
-      0.3675302267074585,
-      0.37116730213165283,
-      0.37062546610832214,
-      0.37135010957717896,
-      0.3697737455368042,
-      0.3697727620601654,
-      0.3714151084423065,
-      0.37123236060142517,
-      0.3738124966621399,
-      0.3752833604812622,
-      0.37488511204719543,
-      0.3777393102645874,
-      0.37878528237342834,
-      0.3812442421913147,
-      0.38351762294769287,
-      0.38525545597076416,
-      0.3856000006198883,
-      0.3862100839614868,
-      0.3856704533100128,
-      0.38774314522743225,
-      0.38804271817207336,
-      0.3885641396045685,
-      0.38839849829673767,
-      0.3905911147594452,
-      0.39113733172416687,
-      0.39186158776283264,
-      0.39222121238708496,
-      0.39306116104125977,
-      0.3945384621620178,
-      0.39511334896087646,
-      0.39469024538993835,
-      0.39563339948654175,
-      0.39637646079063416,
-      0.3954394459724426,
-      0.39604392647743225,
-      0.3968430161476135,
-      0.3972108066082001,
-      0.39802664518356323,
-      0.4002288281917572,
-      0.4006965458393097,
-      0.40092286467552185,
-      0.4012592136859894,
-      0.4014245271682739,
-      0.4025733470916748,
-      0.4031515419483185,
-      0.4036944806575775,
-      0.40446820855140686,
-      0.40457797050476074,
-      0.4052768647670746,
-      0.4059322774410248,
-      0.4069337546825409,
-      0.4073435664176941,
-      0.40719661116600037,
-      0.40762969851493835,
-      0.40837234258651733,
-      0.40862128138542175,
-      0.4101361334323883,
-      0.41058194637298584,
-      0.4111683964729309,
-      0.4115133285522461,
-      0.41178953647613525,
-      0.41227221488952637
     ],
     "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,
-      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,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      8e-05,
-      7.932818532818534e-05,
-      7.816988416988418e-05,
-      7.701158301158302e-05,
-      7.585328185328185e-05,
-      7.469498069498071e-05,
-      7.353667953667954e-05,
-      7.237837837837838e-05,
-      7.122007722007721e-05,
-      7.006177606177606e-05,
-      6.890347490347492e-05,
-      6.774517374517375e-05,
-      6.65868725868726e-05,
-      6.542857142857144e-05,
-      6.427027027027027e-05,
-      6.311196911196911e-05,
-      6.121235521235521e-05,
-      6.0054054054054064e-05,
-      5.8895752895752895e-05,
-      5.773745173745175e-05,
-      5.6579150579150584e-05,
-      5.542084942084943e-05,
-      5.426254826254825e-05,
-      5.310424710424711e-05,
-      5.194594594594594e-05,
-      5.0787644787644786e-05,
-      4.9629343629343644e-05,
-      4.8471042471042475e-05,
-      4.7312741312741326e-05,
-      4.615444015444014e-05,
-      4.4996138996139e-05,
-      4.309652509652511e-05,
-      4.1938223938223946e-05,
-      4.07799227799228e-05,
-      3.962162162162162e-05,
-      3.846332046332047e-05,
-      3.73050193050193e-05,
-      3.6146718146718155e-05,
-      3.4988416988416986e-05,
-      3.383011583011584e-05,
-      3.267181467181467e-05,
-      3.151351351351352e-05,
-      3.0355212355212367e-05,
-      2.9196911196911198e-05,
-      2.8038610038610046e-05,
-      2.6880308880308876e-05,
-      2.4980694980694983e-05,
-      2.3822393822393838e-05,
-      2.266409266409267e-05,
-      2.1505791505791517e-05,
-      2.0347490347490348e-05,
-      1.9189189189189195e-05,
-      1.8030888030888026e-05,
-      1.6872586872586878e-05,
-      1.571428571428571e-05,
-      1.455598455598456e-05,
-      1.3397683397683389e-05,
-      1.223938223938224e-05,
-      1.1081081081081092e-05,
-      9.92277992277992e-06,
-      8.764478764478772e-06
     ],
     "emb_lr": [],
     "eval_step": [
@@ -1231,520 +1231,520 @@
       7788
     ],
     "eval_accuracy": [
-      0.03,
-      0.92,
       0.95,
       0.99,
       1.0,
-      1.0,
-      1.0,
-      1.0,
-      1.0,
       1.0
     ]
   },
-  "final_accuracy": 1.0,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
       "K": null,
       "mode": "sft",
       "n_digits": 6,
-      "n_per_split": 50
     },
     "splits": {
       "add_S0": {
-        "full_accuracy": 0.66,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.9457627118644067,
-            "count": 295
           },
           "SS": {
-            "accuracy": 0.9272727272727272,
-            "count": 55
           }
         }
       },
       "add_S1": {
-        "full_accuracy": 0.78,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.9761904761904762,
-            "count": 126
           },
           "SC": {
-            "accuracy": 0.9367088607594937,
-            "count": 79
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 21
           },
           "UC": {
-            "accuracy": 0.9758064516129032,
-            "count": 124
           }
         }
       },
       "add_S2": {
-        "full_accuracy": 0.42,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.92,
-            "count": 75
           },
           "SC": {
-            "accuracy": 0.9193548387096774,
-            "count": 62
           },
           "SS": {
-            "accuracy": 0.8205128205128205,
-            "count": 39
           },
           "UC": {
-            "accuracy": 0.8378378378378378,
-            "count": 111
           },
           "US": {
-            "accuracy": 0.9047619047619048,
-            "count": 63
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 0.34,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.95,
-            "count": 60
           },
           "SC": {
-            "accuracy": 0.8947368421052632,
-            "count": 57
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 19
           },
           "UC": {
-            "accuracy": 0.7884615384615384,
-            "count": 104
           },
           "US": {
-            "accuracy": 0.8545454545454545,
-            "count": 110
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 0.36,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 48
           },
           "SC": {
-            "accuracy": 0.9807692307692307,
-            "count": 52
           },
           "SS": {
-            "accuracy": 0.8571428571428571,
-            "count": 7
           },
           "UC": {
-            "accuracy": 0.7303370786516854,
-            "count": 89
           },
           "US": {
-            "accuracy": 0.6688311688311688,
-            "count": 154
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.18,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
           },
           "SC": {
-            "accuracy": 0.98,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.44,
-            "count": 50
           },
           "US": {
-            "accuracy": 0.375,
-            "count": 200
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.3,
-        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.34,
-            "count": 50
           },
           "US": {
-            "accuracy": 0.452,
-            "count": 250
           }
         }
       },
       "add_random": {
-        "full_accuracy": 0.795,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.9791183294663574,
-            "count": 431
           },
           "SC": {
-            "accuracy": 0.9778481012658228,
-            "count": 316
           },
           "SS": {
-            "accuracy": 0.8717948717948718,
-            "count": 39
           },
           "UC": {
-            "accuracy": 0.95,
-            "count": 560
           },
           "US": {
-            "accuracy": 0.9259259259259259,
-            "count": 54
           }
         }
       },
       "add_C3": {
-        "full_accuracy": 0.58,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 150
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.8365384615384616,
-            "count": 104
           },
           "US": {
-            "accuracy": 0.8260869565217391,
-            "count": 46
           }
         }
       },
       "add_C4": {
-        "full_accuracy": 0.42,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.99,
-            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.8130081300813008,
-            "count": 123
           },
           "US": {
-            "accuracy": 0.6883116883116883,
-            "count": 77
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.32,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
           },
           "SC": {
-            "accuracy": 0.98,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.7792207792207793,
-            "count": 154
           },
           "US": {
-            "accuracy": 0.8541666666666666,
-            "count": 96
           }
         }
       },
       "add_C6": {
-        "full_accuracy": 0.28,
-        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 0.7692307692307693,
-            "count": 182
           },
           "US": {
-            "accuracy": 0.7711864406779662,
-            "count": 118
           }
         }
       },
       "sub_M0": {
-        "full_accuracy": 0.64,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.935374149659864,
-            "count": 294
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
           }
         }
       },
       "sub_M1": {
-        "full_accuracy": 0.84,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.958041958041958,
-            "count": 143
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 69
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 15
           },
           "UB": {
-            "accuracy": 0.975609756097561,
-            "count": 123
           }
         }
       },
       "sub_M2": {
-        "full_accuracy": 0.36,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.9351851851851852,
-            "count": 108
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 52
           },
           "ME": {
-            "accuracy": 0.9615384615384616,
-            "count": 52
           },
           "UB": {
-            "accuracy": 0.6551724137931034,
-            "count": 87
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 51
           }
         }
       },
       "sub_M3": {
-        "full_accuracy": 0.12,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.9893617021276596,
-            "count": 94
           },
           "MB": {
-            "accuracy": 0.9803921568627451,
-            "count": 51
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 25
           },
           "UB": {
-            "accuracy": 0.6410256410256411,
-            "count": 78
           },
           "UD": {
-            "accuracy": 0.6078431372549019,
-            "count": 102
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 0.06,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.99,
-            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 0.66,
-            "count": 50
           },
           "UD": {
-            "accuracy": 0.34,
-            "count": 150
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.2,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 0.82,
-            "count": 50
           },
           "UD": {
-            "accuracy": 0.38,
-            "count": 200
           }
         }
       },
       "sub_random": {
-        "full_accuracy": 0.775,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.9693877551020408,
-            "count": 588
           },
           "MB": {
-            "accuracy": 0.9925373134328358,
-            "count": 268
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 60
           },
           "UB": {
-            "accuracy": 0.9328859060402684,
-            "count": 447
           },
           "UD": {
-            "accuracy": 0.8918918918918919,
-            "count": 37
           }
         }
       },
       "sub_B3": {
-        "full_accuracy": 0.56,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.9933333333333333,
-            "count": 150
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 0.8317757009345794,
-            "count": 107
           },
           "UD": {
-            "accuracy": 0.8372093023255814,
-            "count": 43
           }
         }
       },
       "sub_B4": {
-        "full_accuracy": 0.34,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 0.8157894736842105,
-            "count": 114
           },
           "UD": {
-            "accuracy": 0.6046511627906976,
-            "count": 86
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.28,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 0.7843137254901961,
-            "count": 153
           },
           "UD": {
-            "accuracy": 0.6597938144329897,
-            "count": 97
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.5107142857142857,
-      "total_examples": 1400,
       "n_splits": 22
     }
   },
@@ -1754,160 +1754,160 @@
       "K": 4,
       "mode": "sorl",
       "n_digits": 6,
-      "n_per_split": 50
     },
     "splits": {
       "add_S0": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 295
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 55
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 126
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 79
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 21
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 124
           }
         }
       },
       "add_S2": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 75
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 62
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 111
           },
           "US": {
             "accuracy": 1.0,
-            "count": 63
           }
         }
       },
       "add_S3": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 60
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 57
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 19
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 104
           },
           "US": {
             "accuracy": 1.0,
-            "count": 110
           }
         }
       },
       "add_S4": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 48
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 52
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 7
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 89
           },
           "US": {
             "accuracy": 1.0,
-            "count": 154
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 50
           },
           "US": {
             "accuracy": 1.0,
-            "count": 200
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 50
           },
           "US": {
             "accuracy": 1.0,
-            "count": 250
           }
         }
       },
@@ -1917,239 +1917,239 @@
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 431
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 316
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 560
           },
           "US": {
             "accuracy": 1.0,
-            "count": 54
           }
         }
       },
       "add_C3": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 150
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 104
           },
           "US": {
             "accuracy": 1.0,
-            "count": 46
           }
         }
       },
       "add_C4": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 123
           },
           "US": {
             "accuracy": 1.0,
-            "count": 77
           }
         }
       },
       "add_C5": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 154
           },
           "US": {
             "accuracy": 1.0,
-            "count": 96
           }
         }
       },
       "add_C6": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 182
           },
           "US": {
             "accuracy": 1.0,
-            "count": 118
           }
         }
       },
       "sub_M0": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 294
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 143
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 69
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 15
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 123
           }
         }
       },
       "sub_M2": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 108
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 52
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 52
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 87
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 51
           }
         }
       },
       "sub_M3": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 94
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 51
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 25
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 78
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 102
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 50
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 150
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 50
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 200
           }
         }
       },
@@ -2159,99 +2159,99 @@
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 588
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 268
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 60
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 447
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 37
           }
         }
       },
       "sub_B3": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 150
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 107
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 43
           }
         }
       },
       "sub_B4": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 114
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 86
           }
         }
       },
       "sub_B5": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 153
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 97
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 1.0,
-      "total_examples": 1400,
       "n_splits": 22
     }
   },
-  "sorl_overall_accuracy": 1.0,
-  "sft_overall_accuracy": 0.5107142857142857
 }

       7788
     ],
     "loss": [
+      14.632190704345703,
+      9.926382064819336,
+      6.634365081787109,
+      4.12091064453125,
+      2.3600664138793945,
+      2.5965561866760254,
+      2.2668251991271973,
+      2.4945991039276123,
+      1.87929368019104,
+      1.770228624343872,
+      1.4810309410095215,
+      1.115319013595581,
+      1.2603116035461426,
+      1.118563175201416,
+      0.7639645338058472,
+      0.5783902406692505,
+      -0.2102416753768921,
+      -1.0637309551239014,
+      -2.982658863067627,
+      -5.1272735595703125,
+      -6.820440769195557,
+      -8.708982467651367,
+      -9.34764575958252,
+      -9.479778289794922,
+      -9.714715003967285,
+      -9.847146987915039,
+      -10.891656875610352,
+      -10.567233085632324,
+      -10.487039566040039,
+      -10.561330795288086,
+      -11.278342247009277,
+      -11.53654956817627,
+      -11.39672565460205,
+      -12.42699146270752,
+      -12.283132553100586,
+      -12.111838340759277,
+      -12.558882713317871,
+      -12.664121627807617,
+      -11.877008438110352,
+      -13.080121040344238,
+      -12.89445972442627,
+      -13.657825469970703,
+      -13.041352272033691,
+      -12.739060401916504,
+      -13.011455535888672,
+      -13.518850326538086,
+      -13.68348217010498,
+      -13.794760704040527,
+      -12.652961730957031,
+      -13.282352447509766,
+      -12.864669799804688,
+      -12.425497055053711,
+      -13.485628128051758,
+      -13.69277286529541,
+      -13.639493942260742,
+      -13.366686820983887,
+      -13.520065307617188,
+      -13.455315589904785,
+      -13.581616401672363,
+      -13.953703880310059,
+      -13.787495613098145,
+      -13.657554626464844,
+      -13.825128555297852,
+      -13.480307579040527,
+      -12.544189453125,
+      -11.929388999938965,
+      -10.827520370483398,
+      -8.075313568115234,
+      -6.391434192657471,
+      -5.907987594604492,
+      -4.740344047546387,
+      -5.025309085845947,
+      -4.2770867347717285,
+      -4.229391574859619,
+      -3.865729570388794,
+      -4.085846424102783,
+      -3.6969783306121826,
+      -3.788874387741089,
+      -3.3547849655151367,
+      -3.1135661602020264,
+      -2.8369789123535156,
+      -2.6105422973632812,
+      -3.406431198120117,
+      -2.858949661254883,
+      -2.1879334449768066,
+      -2.340407133102417,
+      -1.6988248825073242,
+      -1.7582203149795532,
+      -2.0069541931152344,
+      -2.3115696907043457,
+      -1.6048082113265991,
+      -3.4057414531707764,
+      -1.5541704893112183,
+      -1.587241768836975,
+      -1.341800570487976,
+      -1.3259270191192627,
+      -1.2093465328216553,
+      -1.1008269786834717,
+      -0.8763017654418945,
+      -1.4300707578659058,
+      -1.191908597946167,
+      -1.174930453300476,
+      -1.2244458198547363,
+      -0.4350616931915283,
+      -1.2295445203781128,
+      -0.4361853301525116,
+      -0.5482154488563538,
+      -1.4029738903045654,
+      -0.9040775895118713,
+      -0.5001223087310791,
+      -0.3795071244239807,
+      -0.7955602407455444,
+      -0.8618952631950378,
+      -0.6030458807945251,
+      -0.8650072813034058,
+      -0.632997989654541,
+      -0.8306871652603149,
+      -0.6155402660369873,
+      -0.47851070761680603,
+      -0.4381093382835388,
+      -0.7446098327636719,
+      -0.604732871055603,
+      -0.586620569229126,
+      -0.5812572240829468,
+      -0.5657588839530945,
+      -0.40892261266708374,
+      -0.34598445892333984,
+      -0.37908679246902466,
+      -0.6343741416931152,
+      -0.48928260803222656,
+      -0.4414480924606323,
+      -0.5256525874137878,
+      -0.5549529790878296,
+      -0.5228503942489624,
+      -0.43323570489883423,
+      -0.35606762766838074,
+      -0.3718273639678955,
+      -0.3244103789329529,
+      -0.31916630268096924,
+      -0.36536484956741333,
+      -0.5733921527862549,
+      -0.36300128698349,
+      -0.16675259172916412,
+      -0.3644050359725952,
+      -0.3043482303619385,
+      -0.24840393662452698,
+      -0.19103999435901642,
+      -0.1850191354751587,
+      -0.3407224416732788,
+      -0.25857996940612793
     ],
     "base_loss": [
+      9.482434272766113,
+      7.117786884307861,
+      6.046176433563232,
+      4.258971214294434,
+      2.312073230743408,
+      1.998913049697876,
+      1.8657360076904297,
+      1.941304087638855,
+      1.8812999725341797,
+      1.8264588117599487,
+      1.827309250831604,
+      1.793294906616211,
+      1.8274389505386353,
+      1.8220479488372803,
+      1.8286134004592896,
+      1.7788950204849243,
+      1.8032491207122803,
+      1.7887276411056519,
+      2.006676197052002,
+      1.9466220140457153,
+      1.938077688217163,
+      2.0098798274993896,
+      1.9562937021255493,
+      1.7426707744598389,
+      1.8475419282913208,
+      1.8222671747207642,
+      1.7838033437728882,
+      1.7463515996932983,
+      1.697568655014038,
+      1.7143851518630981,
+      1.7419077157974243,
+      1.7188345193862915,
+      1.658122181892395,
+      1.7801390886306763,
+      1.6910394430160522,
+      1.6635606288909912,
+      1.7399853467941284,
+      1.6727386713027954,
+      1.5781800746917725,
+      1.6815193891525269,
+      1.6525733470916748,
+      1.708761215209961,
+      1.6608943939208984,
+      1.6103297472000122,
+      1.623249888420105,
+      1.6843823194503784,
+      1.7061597108840942,
+      1.6853832006454468,
+      1.5944339036941528,
+      1.6551153659820557,
+      1.5582751035690308,
+      1.5246549844741821,
+      1.6163057088851929,
+      1.616803526878357,
+      1.6256102323532104,
+      1.5938681364059448,
+      1.6049398183822632,
+      1.5970059633255005,
+      1.624014973640442,
+      1.625469446182251,
+      1.5937610864639282,
+      1.5731550455093384,
+      1.5809533596038818,
+      1.5393472909927368,
+      1.435717225074768,
+      1.369104266166687,
+      1.2582679986953735,
+      0.9665323495864868,
+      0.7821284532546997,
+      0.7074981331825256,
+      0.5917485356330872,
+      0.6081733107566833,
+      0.5498444437980652,
+      0.5120163559913635,
+      0.4534907341003418,
+      0.4987371265888214,
+      0.4344933032989502,
+      0.4616098999977112,
+      0.3905290961265564,
+      0.36903122067451477,
+      0.3468772768974304,
+      0.3053329885005951,
+      0.39532873034477234,
+      0.33305686712265015,
+      0.2736288011074066,
+      0.27657410502433777,
+      0.21591803431510925,
+      0.20601941645145416,
+      0.23737052083015442,
+      0.27516016364097595,
+      0.20490236580371857,
+      0.4041613042354584,
+      0.1861131340265274,
+      0.19129757583141327,
+      0.1886499673128128,
+      0.18159404397010803,
+      0.14922840893268585,
+      0.13664014637470245,
+      0.10598627477884293,
+      0.1828942596912384,
+      0.1439986377954483,
+      0.14368221163749695,
+      0.14707587659358978,
+      0.07975024729967117,
+      0.14613859355449677,
+      0.06020959094166756,
+      0.07016318291425705,
+      0.1634124517440796,
+      0.1087348610162735,
+      0.082696333527565,
+      0.07368756830692291,
+      0.10071809589862823,
+      0.10751090198755264,
+      0.07716800272464752,
+      0.10648422688245773,
+      0.07751994580030441,
+      0.09940018504858017,
+      0.07651485502719879,
+      0.05863216146826744,
+      0.05505487695336342,
+      0.09117519855499268,
+      0.07521039247512817,
+      0.07184553146362305,
+      0.07176653295755386,
+      0.0699726790189743,
+      0.052560459822416306,
+      0.044058240950107574,
+      0.048487093299627304,
+      0.08156915754079819,
+      0.061087582260370255,
+      0.06119650602340698,
+      0.06394444406032562,
+      0.07004927843809128,
+      0.0659731775522232,
+      0.05378962680697441,
+      0.0477081723511219,
+      0.04793006181716919,
+      0.04166543111205101,
+      0.04095941409468651,
+      0.04708099737763405,
+      0.07109672576189041,
+      0.045931022614240646,
+      0.024872872978448868,
+      0.04558694362640381,
+      0.03894834965467453,
+      0.03414854779839516,
+      0.027124684303998947,
+      0.027090419083833694,
+      0.044619057327508926,
+      0.033715397119522095
     ],
     "info_loss": [
+      -0.12326622009277344,
+      -0.18317556381225586,
+      -0.21192121505737305,
+      -0.1614665985107422,
+      -0.09911632537841797,
+      -0.03676164150238037,
+      -0.05518639087677002,
+      -0.03937363624572754,
+      -0.09468269348144531,
+      -0.09933149814605713,
+      -0.127937912940979,
+      -0.1626497507095337,
+      -0.15077292919158936,
+      -0.16379308700561523,
+      -0.19910144805908203,
+      -0.20818078517913818,
+      -0.2778512239456177,
+      -0.351151704788208,
+      -0.5528711080551147,
+      -0.7527784109115601,
+      -0.9174032211303711,
+      -1.1097712516784668,
+      -1.1687355041503906,
+      -1.156583309173584,
+      -1.1916966438293457,
+      -1.1981340646743774,
+      -1.2985475063323975,
+      -1.2606419324874878,
+      -1.246760606765747,
+      -1.2560533285140991,
+      -1.331277847290039,
+      -1.3520653247833252,
+      -1.3316826820373535,
+      -1.4475840330123901,
+      -1.423694372177124,
+      -1.401245355606079,
+      -1.4543802738189697,
+      -1.456540822982788,
+      -1.3692126274108887,
+      -1.4987409114837646,
+      -1.4760082960128784,
+      -1.5583420991897583,
+      -1.4900978803634644,
+      -1.4563288688659668,
+      -1.4812694787979126,
+      -1.5390629768371582,
+      -1.5584094524383545,
+      -1.56727933883667,
+      -1.4423127174377441,
+      -1.5100535154342651,
+      -1.4591412544250488,
+      -1.4114735126495361,
+      -1.5255085229873657,
+      -1.547141432762146,
+      -1.54146146774292,
+      -1.510321855545044,
+      -1.5257158279418945,
+      -1.5199824571609497,
+      -1.5334408283233643,
+      -1.5722233057022095,
+      -1.5500802993774414,
+      -1.5357627868652344,
+      -1.552409052848816,
+      -1.5149730443954468,
+      -1.40906822681427,
+      -1.341101884841919,
+      -1.2200331687927246,
+      -0.9157827496528625,
+      -0.7277940511703491,
+      -0.6731217503547668,
+      -0.5456058979034424,
+      -0.5737265944480896,
+      -0.49359437823295593,
+      -0.48544979095458984,
+      -0.4402371644973755,
+      -0.467752605676651,
+      -0.42211639881134033,
+      -0.434103399515152,
+      -0.38376203179359436,
+      -0.3583239018917084,
+      -0.3261827230453491,
+      -0.29966747760772705,
+      -0.38840124011039734,
+      -0.32828328013420105,
+      -0.25409969687461853,
+      -0.2704025208950043,
+      -0.1985175758600235,
+      -0.20338541269302368,
+      -0.23216019570827484,
+      -0.26735031604766846,
+      -0.18867464363574982,
+      -0.38847246766090393,
+      -0.182339146733284,
+      -0.18704256415367126,
+      -0.1616264432668686,
+      -0.15969634056091309,
+      -0.14530818164348602,
+      -0.13175378739833832,
+      -0.10473348200321198,
+      -0.171867236495018,
+      -0.1426738202571869,
+      -0.13931553065776825,
+      -0.14417681097984314,
+      -0.05767478048801422,
+      -0.14531880617141724,
+      -0.05537816509604454,
+      -0.06925474852323532,
+      -0.162682443857193,
+      -0.10662886500358582,
+      -0.06461586803197861,
+      -0.05278106778860092,
+      -0.09546349942684174,
+      -0.10324031114578247,
+      -0.07626307010650635,
+      -0.10292515903711319,
+      -0.07728055119514465,
+      -0.09876345098018646,
+      -0.07488290220499039,
+      -0.05830881744623184,
+      -0.05468850955367088,
+      -0.09090279042720795,
+      -0.07504025101661682,
+      -0.07148993760347366,
+      -0.07156261801719666,
+      -0.0697956308722496,
+      -0.05241704359650612,
+      -0.04384676739573479,
+      -0.048265501856803894,
+      -0.07867752760648727,
+      -0.06092023849487305,
+      -0.0558161623775959,
+      -0.06381774693727493,
+      -0.06975483149290085,
+      -0.06581415981054306,
+      -0.053555652499198914,
+      -0.0474986732006073,
+      -0.04777195677161217,
+      -0.04158833995461464,
+      -0.04086068645119667,
+      -0.046969663351774216,
+      -0.07098674029111862,
+      -0.04583430290222168,
+      -0.02477315440773964,
+      -0.045496866106987,
+      -0.03886625915765762,
+      -0.033758021891117096,
+      -0.026911893859505653,
+      -0.026989782229065895,
+      -0.0445425882935524,
+      -0.03364324942231178
     ],
     "abs_loss": [
+      3.8987743854522705,
+      3.835925340652466,
+      3.731152057647705,
+      3.3448033332824707,
+      2.9310805797576904,
+      2.8018741607666016,
+      2.832192897796631,
+      2.8128607273101807,
+      2.843754291534424,
+      2.762322187423706,
+      2.6791672706604004,
+      2.8312129974365234,
+      2.7969703674316406,
+      2.793354034423828,
+      2.753342628479004,
+      2.7262606620788574,
+      2.620635986328125,
+      2.5968563556671143,
+      2.31925106048584,
+      2.0016016960144043,
+      1.9219343662261963,
+      1.7101284265518188,
+      1.7175776958465576,
+      1.6339759826660156,
+      1.8100486993789673,
+      1.4601774215698242,
+      1.461230993270874,
+      1.3801701068878174,
+      1.3301787376403809,
+      1.3739302158355713,
+      1.587360143661499,
+      1.3912031650543213,
+      1.3633630275726318,
+      1.4589648246765137,
+      1.4376378059387207,
+      1.2283248901367188,
+      1.3686951398849487,
+      1.261014461517334,
+      1.1936320066452026,
+      1.2106618881225586,
+      1.1354954242706299,
+      1.202142357826233,
+      1.1697142124176025,
+      1.216151237487793,
+      0.9472143054008484,
+      1.0352338552474976,
+      1.2005023956298828,
+      0.9383926391601562,
+      0.9692818522453308,
+      0.8716940879821777,
+      0.9582070112228394,
+      1.061899185180664,
+      0.9574461579322815,
+      1.0999226570129395,
+      0.8257886171340942,
+      0.9105151891708374,
+      0.7686405181884766,
+      0.7620456218719482,
+      0.8083024621009827,
+      0.7366126775741577,
+      0.64865642786026,
+      0.6970773935317993,
+      0.5407489538192749,
+      0.9302119016647339,
+      0.6019563674926758,
+      0.6286913156509399,
+      0.6697531938552856,
+      0.85337895154953,
+      0.605627715587616,
+      0.6994856595993042,
+      0.6913561820983887,
+      0.7227469086647034,
+      0.8089389204978943,
+      0.679562509059906,
+      0.5462037324905396,
+      0.5866653919219971,
+      0.48199528455734253,
+      0.49292898178100586,
+      0.5071792602539062,
+      0.6359560489654541,
+      0.4345041811466217,
+      0.5750279426574707,
+      0.5141332149505615,
+      0.48265522718429565,
+      0.38371920585632324,
+      0.6727322936058044,
+      0.524191677570343,
+      0.46154582500457764,
+      0.38243192434310913,
+      0.5157103538513184,
+      0.48300936818122864,
+      0.5282471179962158,
+      0.47535738348960876,
+      0.4935920834541321,
+      0.4585495889186859,
+      0.4682696759700775,
+      0.3801548182964325,
+      0.3888419568538666,
+      0.314349889755249,
+      0.40311598777770996,
+      0.393927663564682,
+      0.3618808388710022,
+      0.31192538142204285,
+      0.29459425806999207,
+      0.38845521211624146,
+      0.3338332772254944,
+      0.45690685510635376,
+      0.2870320677757263,
+      0.30454182624816895,
+      0.39305591583251953,
+      0.4107697606086731,
+      0.2857040762901306,
+      0.40057629346847534,
+      0.2520090937614441,
+      0.3365999460220337,
+      0.3793593645095825,
+      0.2781757414340973,
+      0.23415732383728027,
+      0.17023888230323792,
+      0.236487478017807,
+      0.29216936230659485,
+      0.18099826574325562,
+      0.28724172711372375,
+      0.15562580525875092,
+      0.24116644263267517,
+      0.29714399576187134,
+      0.23464536666870117,
+      0.25069287419319153,
+      0.24226970970630646,
+      0.27992311120033264,
+      0.286587655544281,
+      0.17398861050605774,
+      0.28588318824768066,
+      0.2069043666124344,
+      0.18253867328166962,
+      0.24283462762832642,
+      0.26914891600608826,
+      0.22705870866775513,
+      0.1970517933368683,
+      0.17965485155582428,
+      0.2492925226688385,
+      0.1700528860092163,
+      0.16555199027061462,
+      0.22087526321411133,
+      0.22362147271633148,
+      0.23984941840171814,
+      0.23964354395866394,
+      0.17572689056396484,
+      0.18303079903125763,
+      0.20619693398475647
     ],
     "zipf_loss": [
+      5.992541313171387,
+      4.256757736206055,
+      2.3342857360839844,
+      1.142124891281128,
+      0.7460483908653259,
+      0.6850723028182983,
+      0.6697337031364441,
+      0.6657452583312988,
+      0.6604452133178711,
+      0.6608524918556213,
+      0.6651840209960938,
+      0.6654003262519836,
+      0.6609048247337341,
+      0.6551106572151184,
+      0.6510313749313354,
+      0.6086770296096802,
+      0.502957820892334,
+      0.39937275648117065,
+      0.3074512779712677,
+      0.25372886657714844,
+      0.22332003712654114,
+      0.20783749222755432,
+      0.2116575390100479,
+      0.17998555302619934,
+      0.1737050563097,
+      0.1659075915813446,
+      0.16389304399490356,
+      0.15481875836849213,
+      0.14997941255569458,
+      0.14742501080036163,
+      0.13379204273223877,
+      0.1261483132839203,
+      0.1256428062915802,
+      0.12281343340873718,
+      0.1190091073513031,
+      0.11422161012887955,
+      0.10806499421596527,
+      0.1024467945098877,
+      0.1175742894411087,
+      0.10470254719257355,
+      0.09950097650289536,
+      0.09661951661109924,
+      0.08176100254058838,
+      0.09228333085775375,
+      0.0832676887512207,
+      0.08387405425310135,
+      0.07440327852964401,
+      0.09880977869033813,
+      0.07880298793315887,
+      0.07589824497699738,
+      0.07264722883701324,
+      0.05839239060878754,
+      0.057406291365623474,
+      0.051846496760845184,
+      0.06693202257156372,
+      0.051612261682748795,
+      0.055289462208747864,
+      0.07129812985658646,
+      0.04794581979513168,
+      0.06939932703971863,
+      0.05468002334237099,
+      0.057209789752960205,
+      0.0639333575963974,
+      0.03705426678061485,
+      0.05057953670620918,
+      0.04965557903051376,
+      0.04756723344326019,
+      0.030643729493021965,
+      0.043815724551677704,
+      0.04578302800655365,
+      0.05483042076230049,
+      0.03150896355509758,
+      0.028118720278143883,
+      0.04513361677527428,
+      0.028530914336442947,
+      0.034275688230991364,
+      0.04149279743432999,
+      0.04125683009624481,
+      0.04158829152584076,
+      0.03704618662595749,
+      0.034520722925662994,
+      0.023296881467103958,
+      0.030839255079627037,
+      0.042560875415802,
+      0.04106279835104942,
+      0.019770771265029907,
+      0.01801365055143833,
+      0.023459654301404953,
+      0.03903378173708916,
+      0.03520215302705765,
+      0.02873474918305874,
+      0.02199714072048664,
+      0.03557206690311432,
+      0.042527176439762115,
+      0.03995896130800247,
+      0.042615462094545364,
+      0.05649139732122421,
+      0.04118660092353821,
+      0.0336117222905159,
+      0.06539583206176758,
+      0.05143829807639122,
+      0.03835456073284149,
+      0.03905396908521652,
+      0.03247641399502754,
+      0.038659460842609406,
+      0.02400336042046547,
+      0.028478197753429413,
+      0.03173477202653885,
+      0.023022141307592392,
+      0.024034500122070312,
+      0.03353904187679291,
+      0.029786277562379837,
+      0.02293926477432251,
+      0.05721593275666237,
+      0.02410007268190384,
+      0.02435165084898472,
+      0.02972968854010105,
+      0.03335817903280258,
+      0.0289214625954628,
+      0.030072137713432312,
+      0.044025976210832596,
+      0.05235940217971802,
+      0.027709070593118668,
+      0.04703980311751366,
+      0.03810810297727585,
+      0.03297300264239311,
+      0.024960467591881752,
+      0.03001183271408081,
+      0.04660506919026375,
+      0.030839864164590836,
+      0.02685823291540146,
+      0.03118157386779785,
+      0.043957799673080444,
+      0.04862763732671738,
+      0.03027738630771637,
+      0.04692744463682175,
+      0.031047239899635315,
+      0.02710168994963169,
+      0.028775956481695175,
+      0.03928530216217041,
+      0.04044933617115021,
+      0.03240543231368065,
+      0.039550866931676865,
+      0.022889159619808197,
+      0.023003853857517242,
+      0.031042777001857758,
+      0.02698989398777485,
+      0.04021558538079262,
+      0.04178132116794586,
+      0.023517422378063202
     ],
     "denoise_loss": [],
     "ortho_loss": [
+      0.6933014988899231,
+      0.37083250284194946,
+      0.16746486723423004,
+      0.0885070413351059,
+      0.06592030823230743,
+      0.053648799657821655,
+      0.04771490767598152,
+      0.04534199461340904,
+      0.0437336266040802,
+      0.045895759016275406,
+      0.05053376406431198,
+      0.0523027703166008,
+      0.05737421289086342,
+      0.061693500727415085,
+      0.06343631446361542,
+      0.06562791019678116,
+      0.06717279553413391,
+      0.07206566631793976,
+      0.07831596583127975,
+      0.07906866073608398,
+      0.08471731096506119,
+      0.09428071230649948,
+      0.10492909699678421,
+      0.11721401661634445,
+      0.12516844272613525,
+      0.1334349811077118,
+      0.14114034175872803,
+      0.1503930687904358,
+      0.15645001828670502,
+      0.16175302863121033,
+      0.1684732288122177,
+      0.16980968415737152,
+      0.1736874133348465,
+      0.17783427238464355,
+      0.18060865998268127,
+      0.18240122497081757,
+      0.1839177906513214,
+      0.18623745441436768,
+      0.1907666027545929,
+      0.19191065430641174,
+      0.19565843045711517,
+      0.19815172255039215,
+      0.20228537917137146,
+      0.20597098767757416,
+      0.2075631469488144,
+      0.20883804559707642,
+      0.2080991268157959,
+      0.21085670590400696,
+      0.21062801778316498,
+      0.2116369903087616,
+      0.2120591253042221,
+      0.21061359345912933,
+      0.20848363637924194,
+      0.20663155615329742,
+      0.2056088000535965,
+      0.20723474025726318,
+      0.2063807100057602,
+      0.20915858447551727,
+      0.20919714868068695,
+      0.20870265364646912,
+      0.2069307267665863,
+      0.20666652917861938,
+      0.2061752825975418,
+      0.2027379870414734,
+      0.20436672866344452,
+      0.20579716563224792,
+      0.20884613692760468,
+      0.21337348222732544,
+      0.21834859251976013,
+      0.21869564056396484,
+      0.2199230194091797,
+      0.22768796980381012,
+      0.23089125752449036,
+      0.23446932435035706,
+      0.23750188946723938,
+      0.2428743839263916,
+      0.2425122857093811,
+      0.24684441089630127,
+      0.24865762889385223,
+      0.24414007365703583,
+      0.2449251115322113,
+      0.24844960868358612,
+      0.2532559931278229,
+      0.2549189329147339,
+      0.25830599665641785,
+      0.256623238325119,
+      0.25953999161720276,
+      0.2646258771419525,
+      0.26612550020217896,
+      0.26787811517715454,
+      0.2701391279697418,
+      0.27274689078330994,
+      0.274588942527771,
+      0.27538907527923584,
+      0.27801740169525146,
+      0.2798006534576416,
+      0.27908483147621155,
+      0.2781679928302765,
+      0.28107044100761414,
+      0.2800515592098236,
+      0.28154537081718445,
+      0.2851882874965668,
+      0.29015690088272095,
+      0.2916761636734009,
+      0.29237350821495056,
+      0.29165220260620117,
+      0.295254647731781,
+      0.29890671372413635,
+      0.30145129561424255,
+      0.3019912540912628,
+      0.30194753408432007,
+      0.3042812943458557,
+      0.30524149537086487,
+      0.3049549162387848,
+      0.3069272041320801,
+      0.30974724888801575,
+      0.31134963035583496,
+      0.31236183643341064,
+      0.3140103220939636,
+      0.3156379163265228,
+      0.3186745047569275,
+      0.31933704018592834,
+      0.3208130896091461,
+      0.3216458857059479,
+      0.32485339045524597,
+      0.3259000778198242,
+      0.3260437250137329,
+      0.32846739888191223,
+      0.3293222486972809,
+      0.33021095395088196,
+      0.33130118250846863,
+      0.33234986662864685,
+      0.3325884938240051,
+      0.33358755707740784,
+      0.3331451416015625,
+      0.33463388681411743,
+      0.33540332317352295,
+      0.3352293372154236,
+      0.3365360498428345,
+      0.33647823333740234,
+      0.3373063802719116,
+      0.3373261094093323,
+      0.3375466763973236,
+      0.33824118971824646,
+      0.3386029601097107,
+      0.33914104104042053,
+      0.3400459289550781,
+      0.3409873843193054,
+      0.3414129614830017,
+      0.3418258726596832
     ],
     "lr": [
+      8.376068376068378e-06,
+      1.6923076923076924e-05,
+      2.5470085470085475e-05,
+      3.401709401709402e-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,
+      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,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      3.993593461639863e-05,
+      3.934273662008964e-05,
+      3.874953862378065e-05,
+      3.815634062747166e-05,
+      3.7563142631162665e-05,
+      3.696994463485368e-05,
+      3.637674663854469e-05,
+      3.5783548642235693e-05,
+      3.519035064592671e-05,
+      3.459715264961771e-05,
+      3.400395465330873e-05,
+      3.341075665699973e-05,
+      3.281755866069075e-05,
+      3.2224360664381764e-05,
+      3.1251515950435014e-05,
+      3.065831795412602e-05,
+      3.0065119957817037e-05,
+      2.947192196150804e-05,
+      2.887872396519905e-05,
+      2.8285525968890065e-05,
+      2.769232797258107e-05,
+      2.709912997627208e-05,
+      2.6505931979963087e-05,
+      2.5912733983654104e-05,
+      2.5319535987345117e-05,
+      2.472633799103612e-05,
+      2.4133139994727132e-05,
+      2.3539941998418135e-05,
+      2.2946744002109148e-05,
+      2.1973899288162408e-05,
+      2.138070129185341e-05,
+      2.0787503295544424e-05,
+      2.019430529923544e-05,
+      1.9601107302926443e-05,
+      1.900790930661746e-05,
+      1.8414711310308462e-05,
+      1.7821513313999475e-05,
+      1.7228315317690488e-05,
+      1.663511732138149e-05,
+      1.6041919325072507e-05,
+      1.544872132876351e-05,
+      1.4855523332454524e-05,
+      1.4262325336145537e-05,
+      1.366912733983654e-05,
+      1.2696282625889797e-05,
+      1.2103084629580812e-05,
+      1.1509886633271816e-05,
+      1.0916688636962829e-05,
+      1.0323490640653833e-05,
+      9.730292644344846e-06,
+      9.137094648035861e-06,
+      8.543896651726864e-06,
+      7.950698655417878e-06,
+      7.357500659108881e-06,
+      6.764302662799895e-06,
+      6.171104666490909e-06,
+      5.577906670181913e-06,
+      4.984708673872927e-06,
+      4.391510677563931e-06
     ],
     "emb_lr": [],
     "eval_step": [
       7788
     ],
     "eval_accuracy": [
+      0.0,
+      0.18,
+      0.74,
+      0.94,
+      0.99,
       0.95,
       0.99,
       1.0,
+      0.99,
       1.0
     ]
   },
+  "final_accuracy": 0.9983333333333333,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
       "K": null,
       "mode": "sft",
       "n_digits": 6,
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
+        "full_accuracy": 0.98,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
+            "accuracy": 0.9983471074380166,
+            "count": 605
           },
           "SS": {
+            "accuracy": 0.9894736842105263,
+            "count": 95
           }
         }
       },
       "add_S1": {
+        "full_accuracy": 0.98,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
+            "accuracy": 0.9950980392156863,
+            "count": 204
           },
           "SC": {
+            "accuracy": 0.9940828402366864,
+            "count": 169
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 31
           },
           "UC": {
+            "accuracy": 1.0,
+            "count": 296
           }
         }
       },
       "add_S2": {
+        "full_accuracy": 0.96,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
+            "accuracy": 1.0,
+            "count": 163
           },
           "SC": {
+            "accuracy": 0.9769230769230769,
+            "count": 130
           },
           "SS": {
+            "accuracy": 0.9770114942528736,
+            "count": 87
           },
           "UC": {
+            "accuracy": 1.0,
+            "count": 203
           },
           "US": {
+            "accuracy": 1.0,
+            "count": 117
           }
         }
       },
       "add_S3": {
+        "full_accuracy": 0.76,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
+            "accuracy": 1.0,
+            "count": 121
           },
           "SC": {
+            "accuracy": 1.0,
+            "count": 121
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 49
           },
           "UC": {
+            "accuracy": 0.8709677419354839,
+            "count": 186
           },
           "US": {
+            "accuracy": 1.0,
+            "count": 223
           }
         }
       },
       "add_S4": {
+        "full_accuracy": 0.55,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 104
           },
           "SC": {
+            "accuracy": 1.0,
+            "count": 106
           },
           "SS": {
+            "accuracy": 1.0,
+            "count": 23
           },
           "UC": {
+            "accuracy": 0.8,
+            "count": 160
           },
           "US": {
+            "accuracy": 0.9022801302931596,
+            "count": 307
           }
         }
       },
       "add_S5": {
+        "full_accuracy": 0.32,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 100
           },
           "SC": {
+            "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.49,
+            "count": 100
           },
           "US": {
+            "accuracy": 0.6975,
+            "count": 400
           }
         }
       },
       "add_S6": {
+        "full_accuracy": 0.61,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.67,
+            "count": 100
           },
           "US": {
+            "accuracy": 0.782,
+            "count": 500
           }
         }
       },
       "add_random": {
+        "full_accuracy": 0.975,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
+            "accuracy": 1.0,
+            "count": 447
           },
           "SC": {
+            "accuracy": 0.996875,
+            "count": 320
           },
           "SS": {
+            "accuracy": 1.0,
+            "count": 56
           },
           "UC": {
+            "accuracy": 0.9924385633270322,
+            "count": 529
           },
           "US": {
+            "accuracy": 1.0,
+            "count": 48
           }
         }
       },
       "add_C3": {
+        "full_accuracy": 0.83,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 300
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.9119170984455959,
+            "count": 193
           },
           "US": {
+            "accuracy": 0.9626168224299065,
+            "count": 107
           }
         }
       },
       "add_C4": {
+        "full_accuracy": 0.8,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
+            "accuracy": 1.0,
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.92578125,
+            "count": 256
           },
           "US": {
+            "accuracy": 0.9513888888888888,
+            "count": 144
           }
         }
       },
       "add_C5": {
+        "full_accuracy": 0.84,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 100
           },
           "SC": {
+            "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.9575163398692811,
+            "count": 306
           },
           "US": {
+            "accuracy": 0.9639175257731959,
+            "count": 194
           }
         }
       },
       "add_C6": {
+        "full_accuracy": 0.8,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.9480874316939891,
+            "count": 366
           },
           "US": {
+            "accuracy": 0.9914529914529915,
+            "count": 234
           }
         }
       },
       "sub_M0": {
+        "full_accuracy": 0.97,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 0.9950083194675541,
+            "count": 601
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 99
           }
         }
       },
       "sub_M1": {
+        "full_accuracy": 0.98,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 0.996415770609319,
+            "count": 279
           },
           "MB": {
+            "accuracy": 0.993103448275862,
+            "count": 145
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 24
           },
           "UB": {
+            "accuracy": 1.0,
+            "count": 252
           }
         }
       },
       "sub_M2": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 1.0,
+            "count": 213
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 113
           },
           "ME": {
+            "accuracy": 1.0,
+            "count": 85
           },
           "UB": {
+            "accuracy": 1.0,
+            "count": 181
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 108
           }
         }
       },
       "sub_M3": {
+        "full_accuracy": 0.54,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 1.0,
+            "count": 179
           },
           "MB": {
+            "accuracy": 1.0,
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 56
           },
           "UB": {
+            "accuracy": 0.6912751677852349,
+            "count": 149
           },
           "UD": {
+            "accuracy": 1.0,
+            "count": 213
           }
         }
       },
       "sub_M4": {
+        "full_accuracy": 0.16,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 1.0,
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.38,
+            "count": 100
           },
           "UD": {
+            "accuracy": 0.7733333333333333,
+            "count": 300
           }
         }
       },
       "sub_M5": {
+        "full_accuracy": 0.02,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.28,
+            "count": 100
           },
           "UD": {
+            "accuracy": 0.5325,
+            "count": 400
           }
         }
       },
       "sub_random": {
+        "full_accuracy": 0.985,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
+            "accuracy": 0.9983333333333333,
+            "count": 600
           },
           "MB": {
+            "accuracy": 0.9962546816479401,
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 53
           },
           "UB": {
+            "accuracy": 0.9977220956719818,
+            "count": 439
           },
           "UD": {
+            "accuracy": 1.0,
+            "count": 41
           }
         }
       },
       "sub_B3": {
+        "full_accuracy": 0.94,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
+            "accuracy": 1.0,
+            "count": 300
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.9695431472081218,
+            "count": 197
           },
           "UD": {
+            "accuracy": 1.0,
+            "count": 103
           }
         }
       },
       "sub_B4": {
+        "full_accuracy": 0.73,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.8906882591093117,
+            "count": 247
           },
           "UD": {
+            "accuracy": 0.9477124183006536,
+            "count": 153
           }
         }
       },
       "sub_B5": {
+        "full_accuracy": 0.69,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.9194630872483222,
+            "count": 298
           },
           "UD": {
+            "accuracy": 0.8960396039603961,
+            "count": 202
           }
         }
       }
     },
     "summary": {
+      "overall_accuracy": 0.7658333333333334,
+      "total_examples": 2400,
       "n_splits": 22
     }
   },
       "K": 4,
       "mode": "sorl",
       "n_digits": 6,
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 605
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 95
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 204
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 169
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 31
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 296
           }
         }
       },
       "add_S2": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 163
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 130
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 87
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 203
           },
           "US": {
             "accuracy": 1.0,
+            "count": 117
           }
         }
       },
       "add_S3": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 121
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 121
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 49
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 186
           },
           "US": {
             "accuracy": 1.0,
+            "count": 223
           }
         }
       },
       "add_S4": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 104
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 106
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 23
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 160
           },
           "US": {
             "accuracy": 1.0,
+            "count": 307
           }
         }
       },
       "add_S5": {
+        "full_accuracy": 0.99,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.99,
+            "count": 100
           },
           "US": {
             "accuracy": 1.0,
+            "count": 400
           }
         }
       },
       "add_S6": {
+        "full_accuracy": 0.99,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
+            "accuracy": 0.99,
+            "count": 100
           },
           "US": {
             "accuracy": 1.0,
+            "count": 500
           }
         }
       },
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 447
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 320
           },
           "SS": {
             "accuracy": 1.0,
+            "count": 56
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 529
           },
           "US": {
             "accuracy": 1.0,
+            "count": 48
           }
         }
       },
       "add_C3": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 300
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 193
           },
           "US": {
             "accuracy": 1.0,
+            "count": 107
           }
         }
       },
       "add_C4": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 256
           },
           "US": {
             "accuracy": 1.0,
+            "count": 144
           }
         }
       },
       "add_C5": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 306
           },
           "US": {
             "accuracy": 1.0,
+            "count": 194
           }
         }
       },
       "add_C6": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UC": {
             "accuracy": 1.0,
+            "count": 366
           },
           "US": {
             "accuracy": 1.0,
+            "count": 234
           }
         }
       },
       "sub_M0": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 601
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 99
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 279
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 145
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 24
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 252
           }
         }
       },
       "sub_M2": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 213
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 113
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 85
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 181
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 108
           }
         }
       },
       "sub_M3": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 179
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 56
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 149
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 213
           }
         }
       },
       "sub_M4": {
+        "full_accuracy": 0.99,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.99,
+            "count": 100
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 300
           }
         }
       },
       "sub_M5": {
+        "full_accuracy": 0.99,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
+            "accuracy": 0.99,
+            "count": 100
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 400
           }
         }
       },
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 600
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
+            "count": 53
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 439
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 41
           }
         }
       },
       "sub_B3": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 300
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 197
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 103
           }
         }
       },
       "sub_B4": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 247
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 153
           }
         }
       },
       "sub_B5": {
         "full_accuracy": 1.0,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
+            "count": 100
           },
           "UB": {
             "accuracy": 1.0,
+            "count": 298
           },
           "UD": {
             "accuracy": 1.0,
+            "count": 202
           }
         }
       }
     },
     "summary": {
+      "overall_accuracy": 0.9983333333333333,
+      "total_examples": 2400,
       "n_splits": 22
     }
   },
+  "sorl_overall_accuracy": 0.9983333333333333,
+  "sft_overall_accuracy": 0.7658333333333334
 }

add_sub_sorl_v1_abs50_50K/model.safetensors CHANGED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3807a271457535e90facd4a2c2b7452fb1470e8345fa33b069288d979eff66d5
 size 650466940

 version https://git-lfs.github.com/spec/v1
+oid sha256:6342cf5867d670c96b8fb7972956a9553f7a1f0ec1e2656c505d55780ccbe665
 size 650466940

add_sub_sorl_v1_abs50_50K/train_config.json CHANGED Viewed

@@ -1,35 +1,84 @@
 {
-  "mode": "sorl",
-  "ops": "add_sub",
-  "n_digits": 6,
-  "n_layer": 2,
-  "n_head": 3,
-  "n_embd": 510,
-  "abs_vocab": 50,
   "K": 4,
   "alpha_info_gain": 10.0,
   "alpha_abs": 0.1,
   "alpha_soft_zipf": 1.0,
   "batch_size": 64,
   "num_epochs": 10,
-  "dataset_size": 50000,
-  "lr": 8e-05,
   "output_dir": "ckpt/sweep/as_sorl_abs50_K4_50K",
   "device": "cuda",
   "push_to_hub": true,
   "no_wandb": false,
   "n_params": 162540062,
   "run_name": "add_sub_sorl_v1_abs50_50K",
-  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
-  "timestamp": "2026-04-12T03:52:52.696357+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": "d8644t69",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/d8644t69",
-  "final_accuracy": 1.0,
-  "sft_accuracy": 0.5107142857142857,
   "eval_method": "ArithmeticEvaluator"
 }

 {
+  "num_rollouts": 4,
   "K": 4,
+  "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": 234,
+  "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,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 781,
+  "save_every": 999999,
+  "eval_samples": 100,
   "output_dir": "ckpt/sweep/as_sorl_abs50_K4_50K",
+  "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": 50,
+  "dataset_size": 50000,
+  "mode": "sorl",
   "device": "cuda",
   "push_to_hub": true,
   "no_wandb": false,
   "n_params": 162540062,
   "run_name": "add_sub_sorl_v1_abs50_50K",
+  "git_commit": "f447da529caceac8c7d256cbb2cd185cbc50feac",
+  "timestamp": "2026-04-12T12:20:29.793903+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": "wtwip5r3",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/wtwip5r3",
+  "final_accuracy": 0.9983333333333333,
+  "sft_accuracy": 0.7658333333333334,
   "eval_method": "ArithmeticEvaluator"
 }