Upload add_sub_baseline_25K

add_sub_baseline_25K/metrics.json

@@ -159,478 +159,478 @@
       7800
     ],
     "loss": [
-      11.06562614440918,
-      9.070802688598633,
-      7.564438819885254,
-      7.00793981552124,
-      6.226614952087402,
-      5.854918479919434,
-      4.9685893058776855,
-      4.255834102630615,
-      3.3195507526397705,
-      2.7683334350585938,
-      2.22994065284729,
-      1.9851258993148804,
-      1.8262479305267334,
-      1.8066816329956055,
-      1.725418210029602,
-      1.7170584201812744,
-      1.760280728340149,
-      1.7368391752243042,
-      1.631116509437561,
-      1.628551721572876,
-      1.4732780456542969,
-      1.0680259466171265,
-      0.8227473497390747,
-      0.6230937242507935,
-      0.5450628995895386,
-      0.41445615887641907,
-      0.37112170457839966,
-      0.2718469500541687,
-      0.25747212767601013,
-      0.2155759632587433,
-      0.250948965549469,
-      0.16987545788288116,
-      0.17013859748840332,
-      0.13095267117023468,
-      0.1641116440296173,
-      0.15072695910930634,
-      0.14167438447475433,
-      0.11960385739803314,
-      0.08400659263134003,
-      0.1277170479297638,
-      0.10943807661533356,
-      0.10479181259870529,
-      0.11750847846269608,
-      0.09830878674983978,
-      0.09297787398099899,
-      0.08759058266878128,
-      0.09592296183109283,
-      0.056994277983903885,
-      0.08595649152994156,
-      0.09432648122310638,
-      0.06118571758270264,
-      0.05625522881746292,
-      0.0965944156050682,
-      0.08749499171972275,
-      0.08019021898508072,
-      0.05185790732502937,
-      0.07028406113386154,
-      0.053000323474407196,
-      0.06424856185913086,
-      0.07609923928976059,
-      0.08353769779205322,
-      0.0732322707772255,
-      0.08481856435537338,
-      0.05480723828077316,
-      0.04933084920048714,
-      0.03942330554127693,
-      0.03680780529975891,
-      0.0338483564555645,
-      0.04500473663210869,
-      0.05092630162835121,
-      0.06185285374522209,
-      0.05888064578175545,
-      0.06084391847252846,
-      0.047385044395923615,
-      0.05746296048164368,
-      0.038728680461645126,
-      0.041216280311346054,
-      0.04032302647829056,
-      0.03952217847108841,
-      0.04412711411714554,
-      0.037145402282476425,
-      0.03596339002251625,
-      0.03715207800269127,
-      0.02417995221912861,
-      0.05104166269302368,
-      0.028953881934285164,
-      0.0322723388671875,
-      0.05351385846734047,
-      0.04065471515059471,
-      0.035228464752435684,
-      0.0399099700152874,
-      0.04321296140551567,
-      0.023695966228842735,
-      0.03870895132422447,
-      0.023070303723216057,
-      0.02995055727660656,
-      0.034223996102809906,
-      0.03115018829703331,
-      0.047400325536727905,
-      0.033650610595941544,
-      0.020459134131669998,
-      0.031521618366241455,
-      0.010215552523732185,
-      0.009668267332017422,
-      0.009839163161814213,
-      0.01371306087821722,
-      0.0191287100315094,
-      0.02995881251990795,
-      0.010273347608745098,
-      0.013487428426742554,
-      0.006214354187250137,
-      0.028749624267220497,
-      0.004388689063489437,
-      0.01441959012299776,
-      0.010049402713775635,
-      0.006188374478369951,
-      0.006518733222037554,
-      0.012474223971366882,
-      0.0018296894850209355,
-      0.009708991274237633,
-      0.012771486304700375,
-      0.006753657478839159,
-      0.006519939284771681,
-      0.0015541197499260306,
-      0.0044125900603830814,
-      0.0019812153186649084,
-      0.009496470913290977,
-      0.014241503551602364,
-      0.004722653888165951,
-      0.003749982686713338,
-      0.009157851338386536,
-      0.002904724795371294,
-      0.002242226619273424,
-      0.012742781080305576,
-      0.002705535152927041,
-      0.0015757112996652722,
-      0.004974766168743372,
-      0.007047231774777174,
-      0.004940888378769159,
-      0.007126574404537678,
-      0.004077407065778971,
-      0.009843333624303341,
-      0.008745082654058933,
-      0.0023113691713660955,
-      0.005770131945610046,
-      0.006909515243023634,
-      0.005357807967811823,
-      0.022677551954984665,
-      0.0067167701199650764,
-      0.002013612538576126,
-      0.00514591159299016,
-      0.004963079001754522,
-      0.0025343645829707384,
-      0.012118958868086338,
-      0.012999330647289753,
-      0.0052215722389519215
+      9.4371337890625,
+      7.074652671813965,
+      6.122837543487549,
+      4.6825385093688965,
+      3.0381083488464355,
+      2.0917437076568604,
+      1.882021188735962,
+      1.9029165506362915,
+      1.7700093984603882,
+      1.6935831308364868,
+      1.5817426443099976,
+      1.4419691562652588,
+      1.0461561679840088,
+      0.7932643294334412,
+      0.6623935699462891,
+      0.5456612706184387,
+      0.5513021349906921,
+      0.4767663776874542,
+      0.38498836755752563,
+      0.4133641719818115,
+      0.31893685460090637,
+      0.3751819133758545,
+      0.31682610511779785,
+      0.3113398253917694,
+      0.2859819829463959,
+      0.30192962288856506,
+      0.2594987452030182,
+      0.19506025314331055,
+      0.23191870748996735,
+      0.2392965406179428,
+      0.2253739833831787,
+      0.17677058279514313,
+      0.21856749057769775,
+      0.1249561682343483,
+      0.17712180316448212,
+      0.1823340654373169,
+      0.19931195676326752,
+      0.17808902263641357,
+      0.11950694024562836,
+      0.17866145074367523,
+      0.17124225199222565,
+      0.112669937312603,
+      0.19379594922065735,
+      0.1307276040315628,
+      0.17944271862506866,
+      0.14239387214183807,
+      0.15844760835170746,
+      0.13379694521427155,
+      0.09916692227125168,
+      0.1193651482462883,
+      0.12891770899295807,
+      0.17134933173656464,
+      0.11273209750652313,
+      0.14026163518428802,
+      0.1875033676624298,
+      0.10574714094400406,
+      0.1429772675037384,
+      0.12849411368370056,
+      0.14711573719978333,
+      0.13475127518177032,
+      0.09672373533248901,
+      0.10411882400512695,
+      0.11657597869634628,
+      0.10133056342601776,
+      0.11380661278963089,
+      0.10492556542158127,
+      0.07867716252803802,
+      0.11208102852106094,
+      0.12357942014932632,
+      0.09030698239803314,
+      0.11589111387729645,
+      0.07594970613718033,
+      0.08090318739414215,
+      0.08280433714389801,
+      0.06660598516464233,
+      0.11626514047384262,
+      0.08782225102186203,
+      0.07667553424835205,
+      0.0659191906452179,
+      0.06984346359968185,
+      0.07098687440156937,
+      0.10126902908086777,
+      0.09939603507518768,
+      0.08991016447544098,
+      0.07744383811950684,
+      0.06926203519105911,
+      0.06146344915032387,
+      0.09036505222320557,
+      0.08489241451025009,
+      0.062018830329179764,
+      0.05275828391313553,
+      0.07086634635925293,
+      0.06931327283382416,
+      0.09127948433160782,
+      0.07429345697164536,
+      0.08549024164676666,
+      0.06286706030368805,
+      0.05378304794430733,
+      0.044635117053985596,
+      0.05906063690781593,
+      0.053794749081134796,
+      0.0547708198428154,
+      0.02951294183731079,
+      0.049380701035261154,
+      0.06013629585504532,
+      0.042153794318437576,
+      0.059404365718364716,
+      0.057506877928972244,
+      0.0508505143225193,
+      0.046686530113220215,
+      0.07300496101379395,
+      0.035853564739227295,
+      0.04073333367705345,
+      0.04065943509340286,
+      0.045939598232507706,
+      0.052995141595602036,
+      0.049060072749853134,
+      0.03872013837099075,
+      0.043473295867443085,
+      0.03402497619390488,
+      0.05528164654970169,
+      0.0507243312895298,
+      0.05915408954024315,
+      0.04674220830202103,
+      0.051885757595300674,
+      0.040509335696697235,
+      0.03638586029410362,
+      0.025626471266150475,
+      0.02101641520857811,
+      0.033242858946323395,
+      0.027668897062540054,
+      0.03064912185072899,
+      0.035983264446258545,
+      0.04090739041566849,
+      0.02467496506869793,
+      0.03506144881248474,
+      0.023822205141186714,
+      0.024419376626610756,
+      0.029216334223747253,
+      0.03664829209446907,
+      0.033825285732746124,
+      0.0254127848893404,
+      0.029008716344833374,
+      0.032890621572732925,
+      0.036917805671691895,
+      0.03223109617829323,
+      0.02562890388071537,
+      0.029643472284078598,
+      0.03162688761949539,
+      0.04186885803937912,
+      0.028349589556455612,
+      0.03455352038145065,
+      0.027812888845801353,
+      0.024674389511346817,
+      0.03943091258406639,
+      0.022551996633410454
     ],
     "base_loss": [
-      11.06562614440918,
-      9.070802688598633,
-      7.564438819885254,
-      7.00793981552124,
-      6.226614952087402,
-      5.854918479919434,
-      4.9685893058776855,
-      4.255834102630615,
-      3.3195507526397705,
-      2.7683334350585938,
-      2.22994065284729,
-      1.9851258993148804,
-      1.8262479305267334,
-      1.8066816329956055,
-      1.725418210029602,
-      1.7170584201812744,
-      1.760280728340149,
-      1.7368391752243042,
-      1.631116509437561,
-      1.628551721572876,
-      1.4732780456542969,
-      1.0680259466171265,
-      0.8227473497390747,
-      0.6230937242507935,
-      0.5450628995895386,
-      0.41445615887641907,
-      0.37112170457839966,
-      0.2718469500541687,
-      0.25747212767601013,
-      0.2155759632587433,
-      0.250948965549469,
-      0.16987545788288116,
-      0.17013859748840332,
-      0.13095267117023468,
-      0.1641116440296173,
-      0.15072695910930634,
-      0.14167438447475433,
-      0.11960385739803314,
-      0.08400659263134003,
-      0.1277170479297638,
-      0.10943807661533356,
-      0.10479181259870529,
-      0.11750847846269608,
-      0.09830878674983978,
-      0.09297787398099899,
-      0.08759058266878128,
-      0.09592296183109283,
-      0.056994277983903885,
-      0.08595649152994156,
-      0.09432648122310638,
-      0.06118571758270264,
-      0.05625522881746292,
-      0.0965944156050682,
-      0.08749499171972275,
-      0.08019021898508072,
-      0.05185790732502937,
-      0.07028406113386154,
-      0.053000323474407196,
-      0.06424856185913086,
-      0.07609923928976059,
-      0.08353769779205322,
-      0.0732322707772255,
-      0.08481856435537338,
-      0.05480723828077316,
-      0.04933084920048714,
-      0.03942330554127693,
-      0.03680780529975891,
-      0.0338483564555645,
-      0.04500473663210869,
-      0.05092630162835121,
-      0.06185285374522209,
-      0.05888064578175545,
-      0.06084391847252846,
-      0.047385044395923615,
-      0.05746296048164368,
-      0.038728680461645126,
-      0.041216280311346054,
-      0.04032302647829056,
-      0.03952217847108841,
-      0.04412711411714554,
-      0.037145402282476425,
-      0.03596339002251625,
-      0.03715207800269127,
-      0.02417995221912861,
-      0.05104166269302368,
-      0.028953881934285164,
-      0.0322723388671875,
-      0.05351385846734047,
-      0.04065471515059471,
-      0.035228464752435684,
-      0.0399099700152874,
-      0.04321296140551567,
-      0.023695966228842735,
-      0.03870895132422447,
-      0.023070303723216057,
-      0.02995055727660656,
-      0.034223996102809906,
-      0.03115018829703331,
-      0.047400325536727905,
-      0.033650610595941544,
-      0.020459134131669998,
-      0.031521618366241455,
-      0.010215552523732185,
-      0.009668267332017422,
-      0.009839163161814213,
-      0.01371306087821722,
-      0.0191287100315094,
-      0.02995881251990795,
-      0.010273347608745098,
-      0.013487428426742554,
-      0.006214354187250137,
-      0.028749624267220497,
-      0.004388689063489437,
-      0.01441959012299776,
-      0.010049402713775635,
-      0.006188374478369951,
-      0.006518733222037554,
-      0.012474223971366882,
-      0.0018296894850209355,
-      0.009708991274237633,
-      0.012771486304700375,
-      0.006753657478839159,
-      0.006519939284771681,
-      0.0015541197499260306,
-      0.0044125900603830814,
-      0.0019812153186649084,
-      0.009496470913290977,
-      0.014241503551602364,
-      0.004722653888165951,
-      0.003749982686713338,
-      0.009157851338386536,
-      0.002904724795371294,
-      0.002242226619273424,
-      0.012742781080305576,
-      0.002705535152927041,
-      0.0015757112996652722,
-      0.004974766168743372,
-      0.007047231774777174,
-      0.004940888378769159,
-      0.007126574404537678,
-      0.004077407065778971,
-      0.009843333624303341,
-      0.008745082654058933,
-      0.0023113691713660955,
-      0.005770131945610046,
-      0.006909515243023634,
-      0.005357807967811823,
-      0.022677551954984665,
-      0.0067167701199650764,
-      0.002013612538576126,
-      0.00514591159299016,
-      0.004963079001754522,
-      0.0025343645829707384,
-      0.012118958868086338,
-      0.012999330647289753,
-      0.0052215722389519215
+      9.4371337890625,
+      7.074652671813965,
+      6.122837543487549,
+      4.6825385093688965,
+      3.0381083488464355,
+      2.0917437076568604,
+      1.882021188735962,
+      1.9029165506362915,
+      1.7700093984603882,
+      1.6935831308364868,
+      1.5817426443099976,
+      1.4419691562652588,
+      1.0461561679840088,
+      0.7932643294334412,
+      0.6623935699462891,
+      0.5456612706184387,
+      0.5513021349906921,
+      0.4767663776874542,
+      0.38498836755752563,
+      0.4133641719818115,
+      0.31893685460090637,
+      0.3751819133758545,
+      0.31682610511779785,
+      0.3113398253917694,
+      0.2859819829463959,
+      0.30192962288856506,
+      0.2594987452030182,
+      0.19506025314331055,
+      0.23191870748996735,
+      0.2392965406179428,
+      0.2253739833831787,
+      0.17677058279514313,
+      0.21856749057769775,
+      0.1249561682343483,
+      0.17712180316448212,
+      0.1823340654373169,
+      0.19931195676326752,
+      0.17808902263641357,
+      0.11950694024562836,
+      0.17866145074367523,
+      0.17124225199222565,
+      0.112669937312603,
+      0.19379594922065735,
+      0.1307276040315628,
+      0.17944271862506866,
+      0.14239387214183807,
+      0.15844760835170746,
+      0.13379694521427155,
+      0.09916692227125168,
+      0.1193651482462883,
+      0.12891770899295807,
+      0.17134933173656464,
+      0.11273209750652313,
+      0.14026163518428802,
+      0.1875033676624298,
+      0.10574714094400406,
+      0.1429772675037384,
+      0.12849411368370056,
+      0.14711573719978333,
+      0.13475127518177032,
+      0.09672373533248901,
+      0.10411882400512695,
+      0.11657597869634628,
+      0.10133056342601776,
+      0.11380661278963089,
+      0.10492556542158127,
+      0.07867716252803802,
+      0.11208102852106094,
+      0.12357942014932632,
+      0.09030698239803314,
+      0.11589111387729645,
+      0.07594970613718033,
+      0.08090318739414215,
+      0.08280433714389801,
+      0.06660598516464233,
+      0.11626514047384262,
+      0.08782225102186203,
+      0.07667553424835205,
+      0.0659191906452179,
+      0.06984346359968185,
+      0.07098687440156937,
+      0.10126902908086777,
+      0.09939603507518768,
+      0.08991016447544098,
+      0.07744383811950684,
+      0.06926203519105911,
+      0.06146344915032387,
+      0.09036505222320557,
+      0.08489241451025009,
+      0.062018830329179764,
+      0.05275828391313553,
+      0.07086634635925293,
+      0.06931327283382416,
+      0.09127948433160782,
+      0.07429345697164536,
+      0.08549024164676666,
+      0.06286706030368805,
+      0.05378304794430733,
+      0.044635117053985596,
+      0.05906063690781593,
+      0.053794749081134796,
+      0.0547708198428154,
+      0.02951294183731079,
+      0.049380701035261154,
+      0.06013629585504532,
+      0.042153794318437576,
+      0.059404365718364716,
+      0.057506877928972244,
+      0.0508505143225193,
+      0.046686530113220215,
+      0.07300496101379395,
+      0.035853564739227295,
+      0.04073333367705345,
+      0.04065943509340286,
+      0.045939598232507706,
+      0.052995141595602036,
+      0.049060072749853134,
+      0.03872013837099075,
+      0.043473295867443085,
+      0.03402497619390488,
+      0.05528164654970169,
+      0.0507243312895298,
+      0.05915408954024315,
+      0.04674220830202103,
+      0.051885757595300674,
+      0.040509335696697235,
+      0.03638586029410362,
+      0.025626471266150475,
+      0.02101641520857811,
+      0.033242858946323395,
+      0.027668897062540054,
+      0.03064912185072899,
+      0.035983264446258545,
+      0.04090739041566849,
+      0.02467496506869793,
+      0.03506144881248474,
+      0.023822205141186714,
+      0.024419376626610756,
+      0.029216334223747253,
+      0.03664829209446907,
+      0.033825285732746124,
+      0.0254127848893404,
+      0.029008716344833374,
+      0.032890621572732925,
+      0.036917805671691895,
+      0.03223109617829323,
+      0.02562890388071537,
+      0.029643472284078598,
+      0.03162688761949539,
+      0.04186885803937912,
+      0.028349589556455612,
+      0.03455352038145065,
+      0.027812888845801353,
+      0.024674389511346817,
+      0.03943091258406639,
+      0.022551996633410454
     ],
     "lr": [
-      2.5063938618925837e-06,
-      5.063938618925831e-06,
-      7.62148337595908e-06,
-      1.0179028132992328e-05,
-      1.2736572890025576e-05,
-      1.5294117647058822e-05,
-      1.7851662404092073e-05,
-      2.040920716112532e-05,
-      2.296675191815857e-05,
-      2.5524296675191817e-05,
-      2.8081841432225065e-05,
-      3.0639386189258316e-05,
-      3.3196930946291564e-05,
-      3.575447570332481e-05,
-      3.831202046035806e-05,
-      4.086956521739131e-05,
-      4.3427109974424555e-05,
-      4.598465473145781e-05,
-      4.854219948849105e-05,
-      5.10997442455243e-05,
-      5.365728900255755e-05,
-      5.62148337595908e-05,
-      5.877237851662404e-05,
-      6.13299232736573e-05,
-      6.388746803069055e-05,
-      6.644501278772379e-05,
-      6.900255754475704e-05,
-      7.156010230179029e-05,
-      7.411764705882354e-05,
-      7.667519181585678e-05,
-      7.923273657289003e-05,
-      7.999382181128958e-05,
-      7.996356588945887e-05,
-      7.990811651495726e-05,
-      7.982750864365423e-05,
-      7.97217930916005e-05,
-      7.9591036502993e-05,
-      7.943532130816183e-05,
-      7.925474567160515e-05,
-      7.904942343010533e-05,
-      7.881948402096506e-05,
-      7.856507240040864e-05,
-      7.828634895220009e-05,
-      7.798348938653556e-05,
-      7.765668462927371e-05,
-      7.730614070157413e-05,
-      7.693207859001933e-05,
-      7.653473410730253e-05,
-      7.611435774356888e-05,
-      7.567121450850376e-05,
-      7.520558376426795e-05,
-      7.471775904938474e-05,
-      7.420804789369019e-05,
-      7.367677162446306e-05,
-      7.312426516385672e-05,
-      7.255087681776069e-05,
-      7.195696805622496e-05,
-      7.13429132855854e-05,
-      7.070909961243422e-05,
-      7.005592659958366e-05,
-      6.938380601417765e-05,
-      6.869316156810923e-05,
-      6.798442865090831e-05,
-      6.725805405526735e-05,
-      6.651449569537871e-05,
-      6.575422231826058e-05,
-      6.497771320825402e-05,
-      6.418545788487704e-05,
-      6.337795579422628e-05,
-      6.255571599412105e-05,
-      6.171925683318781e-05,
-      6.086910562408781e-05,
-      6.0005798311093635e-05,
-      5.912987913222422e-05,
-      5.824190027615158e-05,
-      5.734242153409514e-05,
-      5.643200994692358e-05,
-      5.55112394476862e-05,
-      5.458069049979956e-05,
-      5.364094973111714e-05,
-      5.269260956411309e-05,
-      5.1736267842412726e-05,
-      5.077252745390575e-05,
-      4.980199595067928e-05,
-      4.882528516601063e-05,
-      4.784301082866123e-05,
-      4.685579217471466e-05,
-      4.586425155720376e-05,
-      4.4869014053772686e-05,
-      4.387070707262142e-05,
-      4.286995995698098e-05,
-      4.186740358836888e-05,
-      4.0863669988874755e-05,
-      3.985939192272697e-05,
-      3.885520249739142e-05,
-      3.785173476445388e-05,
-      3.684962132053763e-05,
-      3.584949390850793e-05,
-      3.485198301921461e-05,
-      3.385771749402399e-05,
-      3.2867324128390756e-05,
-      3.188142727671938e-05,
-      3.090064845876465e-05,
-      2.9925605967818972e-05,
-      2.895691448093382e-05,
-      2.799518467142088e-05,
-      2.7041022823877087e-05,
-      2.6095030451976512e-05,
-      2.515780391926975e-05,
-      2.4229934063230064e-05,
-      2.3312005822783295e-05,
-      2.240459786955611e-05,
-      2.150828224307534e-05,
-      2.0623623990148315e-05,
-      1.9751180808651272e-05,
-      1.8891502695950898e-05,
-      1.8045131602180072e-05,
-      1.7212601088586823e-05,
-      1.63944359911718e-05,
-      1.559115208982597e-05,
-      1.48032557831777e-05,
-      1.4031243769353617e-05,
-      1.3275602732854923e-05,
-      1.2536809037746398e-05,
-      1.1815328427351398e-05,
-      1.1111615730642416e-05,
-      1.042611457551213e-05,
-      9.759257109105627e-06,
-      9.111463725390388e-06,
-      8.483142800135428e-06,
-      7.87469043346695e-06,
-      7.286490200162668e-06,
-      6.718912907842181e-06,
-      6.1723163632060055e-06,
-      5.647045146470409e-06,
-      5.143430394140439e-06,
-      4.661789590258008e-06,
-      4.202426366256558e-06,
-      3.7656303095486623e-06,
-      3.351676780967026e-06,
-      2.960826741174141e-06,
-      2.5933265861499514e-06,
-      2.2494079918611923e-06,
-      1.929287768210473e-06,
-      1.6331677223569853e-06,
-      1.3612345314951615e-06,
-      1.1136596251714304e-06,
-      8.905990772131879e-07,
-      6.921935073382368e-07,
-      5.185679925066245e-07,
-      3.698319880708301e-07,
-      2.4607925877392135e-07,
-      1.4738781963932191e-07,
-      7.381988678927255e-08,
-      2.5421838223160798e-08,
-      2.2241845803394615e-09
+      8.376068376068378e-06,
+      1.6923076923076924e-05,
+      2.5470085470085475e-05,
+      3.401709401709402e-05,
+      3.99996141174052e-05,
+      3.9992754396617386e-05,
+      3.997732289238075e-05,
+      3.9953326220867826e-05,
+      3.99207746705195e-05,
+      3.987968219763389e-05,
+      3.9830066420382645e-05,
+      3.97719486112573e-05,
+      3.9705353687948734e-05,
+      3.9630310202663935e-05,
+      3.9546850329884316e-05,
+      3.945500985257116e-05,
+      3.9354828146823805e-05,
+      3.924634816499739e-05,
+      3.9129616417287294e-05,
+      3.900468295178809e-05,
+      3.887160133303572e-05,
+      3.8730428619042037e-05,
+      3.858122533683144e-05,
+      3.842405545649026e-05,
+      3.825898636373997e-05,
+      3.808608883104587e-05,
+      3.790543698727386e-05,
+      3.7717108285908e-05,
+      3.752118347184284e-05,
+      3.7317746546764446e-05,
+      3.710688473313514e-05,
+      3.688868843679738e-05,
+      3.666325120821272e-05,
+      3.643066970235256e-05,
+      3.619104363725791e-05,
+      3.5944475751285765e-05,
+      3.569107175906064e-05,
+      3.5430940306149985e-05,
+      3.516419292248301e-05,
+      3.489094397453285e-05,
+      3.461131061628253e-05,
+      3.4325412738995875e-05,
+      3.403337291981479e-05,
+      3.373531636920496e-05,
+      3.343137087727258e-05,
+      3.3121666758975014e-05,
+      3.280633679824903e-05,
+      3.248551619108036e-05,
+      3.21593424875392e-05,
+      3.1827955532806334e-05,
+      3.149149740721537e-05,
+      3.115011236533647e-05,
+      3.080394677412806e-05,
+      3.045314905018279e-05,
+      3.0097869596094755e-05,
+      2.9738260735975154e-05,
+      2.9374476650144227e-05,
+      2.9006673309027263e-05,
+      2.8635008406283132e-05,
+      2.8259641291194015e-05,
+      2.7880732900345262e-05,
+      2.7498445688624725e-05,
+      2.711294355957111e-05,
+      2.6724391795101242e-05,
+      2.633295698464635e-05,
+      2.5938806953727778e-05,
+      2.5542110692002687e-05,
+      2.5143038280810704e-05,
+      2.4741760820252464e-05,
+      2.433845035583137e-05,
+      2.3933279804690064e-05,
+      2.352642288147312e-05,
+      2.3118054023847876e-05,
+      2.2708348317715223e-05,
+      2.2297481422142487e-05,
+      2.188562949405059e-05,
+      2.1472969112687716e-05,
+      2.1059677203921957e-05,
+      2.064593096438528e-05,
+      2.0231907785501493e-05,
+      1.9817785177430605e-05,
+      1.9403740692962372e-05,
+      1.898995185139145e-05,
+      1.8576596062407023e-05,
+      1.816385055002938e-05,
+      1.7751892276626043e-05,
+      1.7340897867040178e-05,
+      1.6931043532863625e-05,
+      1.6522504996887168e-05,
+      1.611545741776037e-05,
+      1.5710075314893218e-05,
+      1.5306532493631954e-05,
+      1.4905001970740918e-05,
+      1.4505655900222602e-05,
+      1.4108665499507574e-05,
+      1.371420097604592e-05,
+      1.332243145433177e-05,
+      1.2933524903392054e-05,
+      1.2547648064770774e-05,
+      1.2164966381039404e-05,
+      1.178564392486436e-05,
+      1.140984332866172e-05,
+      1.1037725714869483e-05,
+      1.066945062686719e-05,
+      1.0305175960572616e-05,
+      9.945057896744699e-06,
+      9.589250834021969e-06,
+      9.237907322724944e-06,
+      8.891177999451028e-06,
+      8.549211522489898e-06,
+      8.212154508087055e-06,
+      7.88015146758299e-06,
+      7.553344745454641e-06,
+      7.231874458286057e-06,
+      6.915878434694157e-06,
+      6.605492156235467e-06,
+      6.30084869931916e-06,
+      6.002078678151244e-06,
+      5.709310188734507e-06,
+      5.422668753947975e-06,
+      5.14227726972974e-06,
+      4.868255952385965e-06,
+      4.600722287048818e-06,
+      4.339790977305362e-06,
+      4.085573896019013e-06,
+      3.838180037364703e-06,
+      3.5977154700981752e-06,
+      3.364283292079631e-06,
+      3.137983586071065e-06,
+      2.9189133768263488e-06,
+      2.707166589492387e-06,
+      2.5028340093392257e-06,
+      2.3060032428363876e-06,
+      2.1167586800920613e-06,
+      1.9351814586713113e-06,
+      1.7613494288088008e-06,
+      1.5953371200309199e-06,
+      1.437215709201667e-06,
+      1.2870529900059636e-06,
+      1.1449133438834802e-06,
+      1.0108577124254482e-06,
+      8.849435712462972e-07,
+      7.67224905341275e-07,
+      6.57752185940721e-07,
+      5.565723488707586e-07,
+      4.637287744298502e-07,
+      3.79261268789719e-07,
+      3.032060469286724e-07,
+      2.3559571710463747e-07,
+      1.7645926687452908e-07,
+      1.2582205066603127e-07,
+      8.370577890698173e-08,
+      5.012850871717989e-08,
+      2.5104636166479735e-08,
+      8.64489010255598e-09,
+      7.563275509769874e-10
     ],
     "eval_step": [
       390,
@@ -677,530 +677,530 @@
       20
     ],
     "eval_accuracy": [
-      0.0064285714285714285,
-      0.007142857142857143,
-      0.085,
-      0.52,
-      0.5935714285714285,
-      0.6385714285714286,
-      0.6514285714285715,
-      0.7364285714285714,
-      0.7307142857142858,
-      0.7328571428571429,
-      0.755,
-      0.7964285714285714,
-      0.8135714285714286,
-      0.89,
-      0.8621428571428571,
-      0.8892857142857142,
-      0.9235714285714286,
-      0.9035714285714286,
-      0.9135714285714286,
-      0.9207142857142857
+      0.0044444444444444444,
+      0.07111111111111111,
+      0.3566666666666667,
+      0.5133333333333333,
+      0.6433333333333333,
+      0.6266666666666667,
+      0.6777777777777778,
+      0.6422222222222222,
+      0.6966666666666667,
+      0.7366666666666667,
+      0.75,
+      0.7633333333333333,
+      0.7722222222222223,
+      0.7922222222222223,
+      0.8388888888888889,
+      0.8188888888888889,
+      0.84,
+      0.8522222222222222,
+      0.8511111111111112,
+      0.8488888888888889
     ]
   },
-  "final_accuracy": 0.9207142857142857,
+  "final_accuracy": 0.8179166666666666,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
       "K": null,
       "mode": "sft",
       "n_digits": 6,
-      "n_per_split": 50
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.95,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 295
+            "accuracy": 0.9917355371900827,
+            "count": 605
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 55
+            "count": 95
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 126
+            "count": 204
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 79
+            "count": 169
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 21
+            "count": 31
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 124
+            "count": 296
           }
         }
       },
       "add_S2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.91,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 75
+            "accuracy": 0.9938650306748467,
+            "count": 163
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 62
+            "accuracy": 0.9923076923076923,
+            "count": 130
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
+            "count": 87
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 111
+            "accuracy": 0.9655172413793104,
+            "count": 203
           },
           "US": {
             "accuracy": 1.0,
-            "count": 63
+            "count": 117
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.87,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 60
+            "count": 121
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 57
+            "count": 121
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 19
+            "count": 49
           },
           "UC": {
-            "accuracy": 1.0,
-            "count": 104
+            "accuracy": 0.9354838709677419,
+            "count": 186
           },
           "US": {
-            "accuracy": 1.0,
-            "count": 110
+            "accuracy": 0.9820627802690582,
+            "count": 223
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 0.92,
-        "n_examples": 50,
+        "full_accuracy": 0.71,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 48
+            "count": 104
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 106
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 7
+            "count": 23
           },
           "UC": {
-            "accuracy": 0.9550561797752809,
-            "count": 89
+            "accuracy": 0.85,
+            "count": 160
           },
           "US": {
-            "accuracy": 0.9935064935064936,
-            "count": 154
+            "accuracy": 0.9641693811074918,
+            "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.54,
-        "n_examples": 50,
+        "full_accuracy": 0.59,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.64,
-            "count": 50
+            "accuracy": 0.72,
+            "count": 100
           },
           "US": {
-            "accuracy": 0.91,
-            "count": 200
+            "accuracy": 0.8775,
+            "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.92,
-        "n_examples": 50,
+        "full_accuracy": 0.74,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.94,
-            "count": 50
+            "accuracy": 0.85,
+            "count": 100
           },
           "US": {
-            "accuracy": 0.98,
-            "count": 250
+            "accuracy": 0.882,
+            "count": 500
           }
         }
       },
       "add_random": {
-        "full_accuracy": 0.995,
+        "full_accuracy": 0.95,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 1.0,
-            "count": 431
+            "accuracy": 0.9865771812080537,
+            "count": 447
           },
           "SC": {
-            "accuracy": 1.0,
-            "count": 316
+            "accuracy": 0.996875,
+            "count": 320
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 39
+            "count": 56
           },
           "UC": {
-            "accuracy": 0.9982142857142857,
-            "count": 560
+            "accuracy": 0.9924385633270322,
+            "count": 529
           },
           "US": {
             "accuracy": 1.0,
-            "count": 54
+            "count": 48
           }
         }
       },
       "add_C3": {
-        "full_accuracy": 0.98,
-        "n_examples": 50,
+        "full_accuracy": 0.84,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 150
+            "count": 300
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.9903846153846154,
-            "count": 104
+            "accuracy": 0.917098445595855,
+            "count": 193
           },
           "US": {
             "accuracy": 1.0,
-            "count": 46
+            "count": 107
           }
         }
       },
       "add_C4": {
         "full_accuracy": 0.9,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.967479674796748,
-            "count": 123
+            "accuracy": 0.9609375,
+            "count": 256
           },
           "US": {
-            "accuracy": 0.974025974025974,
-            "count": 77
+            "accuracy": 1.0,
+            "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.98,
-        "n_examples": 50,
+        "full_accuracy": 0.89,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.9935064935064936,
-            "count": 154
+            "accuracy": 0.9640522875816994,
+            "count": 306
           },
           "US": {
             "accuracy": 1.0,
-            "count": 96
+            "count": 194
           }
         }
       },
       "add_C6": {
-        "full_accuracy": 0.96,
-        "n_examples": 50,
+        "full_accuracy": 0.9,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.989010989010989,
-            "count": 182
+            "accuracy": 0.9754098360655737,
+            "count": 366
           },
           "US": {
-            "accuracy": 0.9915254237288136,
-            "count": 118
+            "accuracy": 0.9957264957264957,
+            "count": 234
           }
         }
       },
       "sub_M0": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.99,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 294
+            "accuracy": 0.9983361064891847,
+            "count": 601
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 99
           }
         }
       },
       "sub_M1": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.98,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 143
+            "accuracy": 0.996415770609319,
+            "count": 279
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 69
+            "accuracy": 0.993103448275862,
+            "count": 145
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 15
+            "count": 24
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 123
+            "count": 252
           }
         }
       },
       "sub_M2": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.92,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 108
+            "accuracy": 0.9953051643192489,
+            "count": 213
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 52
+            "accuracy": 0.9911504424778761,
+            "count": 113
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 52
+            "count": 85
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 87
+            "accuracy": 0.9668508287292817,
+            "count": 181
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 51
+            "count": 108
           }
         }
       },
       "sub_M3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.8,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 94
+            "count": 179
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 51
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 25
+            "count": 56
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 78
+            "accuracy": 0.8657718120805369,
+            "count": 149
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 102
+            "count": 213
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 0.64,
-        "n_examples": 50,
+        "full_accuracy": 0.4,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.68,
-            "count": 50
+            "accuracy": 0.44,
+            "count": 100
           },
           "UD": {
-            "accuracy": 0.9866666666666667,
-            "count": 150
+            "accuracy": 0.97,
+            "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.24,
-        "n_examples": 50,
+        "full_accuracy": 0.08,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.52,
-            "count": 50
+            "accuracy": 0.25,
+            "count": 100
           },
           "UD": {
-            "accuracy": 0.885,
-            "count": 200
+            "accuracy": 0.6875,
+            "count": 400
           }
         }
       },
       "sub_random": {
-        "full_accuracy": 1.0,
+        "full_accuracy": 0.965,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 588
+            "accuracy": 0.9983333333333333,
+            "count": 600
           },
           "MB": {
-            "accuracy": 1.0,
-            "count": 268
+            "accuracy": 0.9925093632958801,
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 60
+            "count": 53
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 447
+            "accuracy": 0.9908883826879271,
+            "count": 439
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 37
+            "count": 41
           }
         }
       },
       "sub_B3": {
-        "full_accuracy": 1.0,
-        "n_examples": 50,
+        "full_accuracy": 0.89,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 1.0,
-            "count": 150
+            "accuracy": 0.9966666666666667,
+            "count": 300
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 1.0,
-            "count": 107
+            "accuracy": 0.949238578680203,
+            "count": 197
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 43
+            "count": 103
           }
         }
       },
       "sub_B4": {
-        "full_accuracy": 0.9,
-        "n_examples": 50,
+        "full_accuracy": 0.74,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.956140350877193,
-            "count": 114
+            "accuracy": 0.8947368421052632,
+            "count": 247
           },
           "UD": {
-            "accuracy": 1.0,
-            "count": 86
+            "accuracy": 0.9934640522875817,
+            "count": 153
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.82,
-        "n_examples": 50,
+        "full_accuracy": 0.7,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.9477124183006536,
-            "count": 153
+            "accuracy": 0.9060402684563759,
+            "count": 298
           },
           "UD": {
-            "accuracy": 0.979381443298969,
-            "count": 97
+            "accuracy": 0.9653465346534653,
+            "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.9207142857142857,
-      "total_examples": 1400,
+      "overall_accuracy": 0.8179166666666666,
+      "total_examples": 2400,
       "n_splits": 22
     }
   }

add_sub_baseline_25K/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2bf525d5b1ad9ce7ab59b46e66c0d4b353f9086c36262ae393e08d901ea70fb2
+oid sha256:795da9d34e3babd627a021a17d4593d73d1e588cccc1e221657909955751d349
 size 650266922

add_sub_baseline_25K/train_config.json

@@ -1,35 +1,84 @@
 {
-  "mode": "baseline",
-  "ops": "add_sub",
-  "n_digits": 6,
-  "n_layer": 2,
-  "n_head": 3,
-  "n_embd": 510,
-  "abs_vocab": 0,
+  "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": 20,
-  "dataset_size": 25000,
-  "lr": 8e-05,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 390,
+  "save_every": 999999,
+  "eval_samples": 100,
   "output_dir": "ckpt/sweep/add_sub_baseline_25K",
+  "eval_K": 4,
+  "alpha_traj": 0.0,
+  "corrupt_method": "shuffle",
+  "corrupt_ratio": 0.3,
+  "alpha_contrastive": 1.0,
+  "gamma_contrastive": 0.5,
+  "alpha_masked_traj": 0.0,
+  "mask_nl_ratio": 0.3,
+  "mask_nl_mode": "fixed",
+  "mask_nl_fixed_id": 0,
+  "use_ste": true,
+  "n_inner": 1,
+  "random_K": null,
+  "strip_suffix": null,
+  "compress_prefix": null,
+  "random_mem_span": null,
+  "warmup_ratio": 0.03,
+  "beta2": 0.999,
+  "seed": 42,
+  "n_digits": 6,
+  "n_layer": 2,
+  "n_head": 3,
+  "n_embd": 510,
+  "ops": "add_sub",
+  "abs_vocab": 0,
+  "dataset_size": 25000,
+  "mode": "baseline",
   "device": "cuda",
   "push_to_hub": true,
   "no_wandb": false,
   "n_params": 162490082,
   "run_name": "add_sub_baseline_25K",
-  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
-  "timestamp": "2026-04-12T01:58:10.719272+00:00",
+  "git_commit": "78d46f8665a87f4b44bd5894bd34f393f2dea51f",
+  "timestamp": "2026-04-12T08:59:11.035906+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": "sft",
-  "wandb_run_id": "hccg8mv4",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/hccg8mv4",
-  "final_accuracy": 0.9207142857142857,
-  "sft_accuracy": 0.9207142857142857,
+  "wandb_run_id": "4f30dkzf",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/4f30dkzf",
+  "final_accuracy": 0.8179166666666666,
+  "sft_accuracy": 0.8179166666666666,
   "eval_method": "ArithmeticEvaluator"
 }