Upload add_sub_sorl_v1_abs30_10K

add_sub_sorl_v1_abs30_10K/metrics.json

@@ -33,229 +33,229 @@
       1563
     ],
     "loss": [
-      12.37744426727295,
-      7.511226177215576,
-      4.2184038162231445,
-      2.3564629554748535,
-      2.5140318870544434,
-      1.9689207077026367,
-      2.304285764694214,
-      2.3624253273010254,
-      1.6747889518737793,
-      1.6878020763397217,
-      1.355285882949829,
-      1.1105128526687622,
-      1.1249760389328003,
-      1.014577865600586,
-      0.7836594581604004,
-      0.9053444266319275,
-      0.783288836479187,
-      0.7557217478752136,
-      0.6148194074630737,
-      0.27221277356147766,
-      0.543100893497467,
-      -0.13505232334136963,
-      -0.5669998526573181,
-      -0.4266250431537628,
-      -1.53482186794281,
-      -2.418619155883789,
-      -2.2528090476989746,
-      -3.1418380737304688,
-      -3.4164059162139893,
-      -4.483319282531738
+      9.84412956237793,
+      5.1068925857543945,
+      2.6231350898742676,
+      2.307675838470459,
+      2.025418281555176,
+      1.0131561756134033,
+      1.7522368431091309,
+      1.42581307888031,
+      1.1289782524108887,
+      0.9756244421005249,
+      0.5843766927719116,
+      0.12120179086923599,
+      0.038071952760219574,
+      -0.09827154129743576,
+      -0.11171114444732666,
+      -0.15556588768959045,
+      -0.08670347929000854,
+      0.03128141164779663,
+      -0.3896399140357971,
+      -0.008183799684047699,
+      -0.1007656455039978,
+      0.3051470220088959,
+      -0.1488836109638214,
+      0.28637364506721497,
+      -0.23199722170829773,
+      0.12318609654903412,
+      -0.2954173684120178,
+      -0.3596929609775543,
+      -0.22620351612567902,
+      -0.7009453177452087
     ],
     "base_loss": [
-      7.905151844024658,
-      5.81351375579834,
-      3.7999191284179688,
-      2.091179609298706,
-      1.9467359781265259,
-      1.8935550451278687,
-      1.8686275482177734,
-      1.8094427585601807,
-      1.896533727645874,
-      1.813745379447937,
-      1.8012498617172241,
-      1.861536979675293,
-      1.8485734462738037,
-      1.8326823711395264,
-      1.8518121242523193,
-      1.7422752380371094,
-      1.865788459777832,
-      1.840556025505066,
-      1.8511160612106323,
-      1.8595082759857178,
-      1.7606910467147827,
-      1.7480539083480835,
-      1.7231824398040771,
-      1.7938270568847656,
-      1.8038008213043213,
-      1.7478877305984497,
-      1.7873140573501587,
-      1.8223601579666138,
-      1.8396693468093872,
-      1.8911950588226318
+      6.9705986976623535,
+      3.7352306842803955,
+      1.9767621755599976,
+      1.8775386810302734,
+      1.9166576862335205,
+      1.8984237909317017,
+      1.8574403524398804,
+      1.801119089126587,
+      1.865805983543396,
+      1.7841968536376953,
+      1.8087198734283447,
+      1.8459980487823486,
+      1.8696975708007812,
+      1.8336594104766846,
+      1.8405402898788452,
+      1.760048747062683,
+      1.8620561361312866,
+      1.8630913496017456,
+      1.8629968166351318,
+      1.8340078592300415,
+      1.7706114053726196,
+      1.712676763534546,
+      1.6838618516921997,
+      1.7604615688323975,
+      1.7645069360733032,
+      1.6746128797531128,
+      1.7230627536773682,
+      1.7214601039886475,
+      1.7510899305343628,
+      1.6944631338119507
     ],
     "info_loss": [
-      -0.1752934455871582,
-      -0.12497377395629883,
-      -0.10779404640197754,
-      -0.09296655654907227,
-      -0.0572664737701416,
-      -0.10382282733917236,
-      -0.06734299659729004,
-      -0.055501341819763184,
-      -0.13261711597442627,
-      -0.12388193607330322,
-      -0.15424871444702148,
-      -0.1853572130203247,
-      -0.1832747459411621,
-      -0.19157755374908447,
-      -0.2160094976425171,
-      -0.19262421131134033,
-      -0.21582400798797607,
-      -0.2133394479751587,
-      -0.22457683086395264,
-      -0.2426363229751587,
-      -0.1686021089553833,
-      -0.2219860553741455,
-      -0.26084792613983154,
-      -0.25028252601623535,
-      -0.36128807067871094,
-      -0.44187653064727783,
-      -0.4287785291671753,
-      -0.5215833187103271,
-      -0.5496792793273926,
-      -0.6622849702835083
+      -0.23186302185058594,
+      -0.016382217407226562,
+      -0.04995226860046387,
+      -0.06860637664794922,
+      -0.09940493106842041,
+      -0.19964098930358887,
+      -0.12048065662384033,
+      -0.14689147472381592,
+      -0.18259012699127197,
+      -0.18501496315002441,
+      -0.16905105113983154,
+      -0.19646573066711426,
+      -0.20380914211273193,
+      -0.21209895610809326,
+      -0.21376097202301025,
+      -0.20788681507110596,
+      -0.21054041385650635,
+      -0.19847512245178223,
+      -0.2401341199874878,
+      -0.19826948642730713,
+      -0.20111238956451416,
+      -0.15373647212982178,
+      -0.1952725648880005,
+      -0.16008174419403076,
+      -0.21182727813720703,
+      -0.16642332077026367,
+      -0.21538293361663818,
+      -0.21970915794372559,
+      -0.2088148593902588,
+      -0.25221550464630127
     ],
     "abs_loss": [
-      3.363813877105713,
-      3.225320816040039,
-      2.8528130054473877,
-      2.6181869506835938,
-      2.622387409210205,
-      2.5087389945983887,
-      2.5558464527130127,
-      2.5691847801208496,
-      2.4980592727661133,
-      2.6305322647094727,
-      2.508139133453369,
-      2.5618228912353516,
-      2.598161220550537,
-      2.5265064239501953,
-      2.4846131801605225,
-      2.5860702991485596,
-      2.4830150604248047,
-      2.454457998275757,
-      2.4279370307922363,
-      2.3325886726379395,
-      1.648212194442749,
-      1.3609596490859985,
-      1.3366903066635132,
-      1.2300528287887573,
-      1.2981481552124023,
-      1.1689049005508423,
-      1.1186243295669556,
-      1.182748556137085,
-      1.1046546697616577,
-      1.1569410562515259
+      3.347839832305908,
+      2.7904562950134277,
+      2.566986083984375,
+      2.5303356647491455,
+      2.5691452026367188,
+      2.56864595413208,
+      2.542234420776367,
+      2.518596649169922,
+      2.5303497314453125,
+      2.397298812866211,
+      1.649463415145874,
+      0.9939525127410889,
+      0.9239900708198547,
+      0.7970927953720093,
+      0.9154700040817261,
+      0.7975423336029053,
+      0.9060695171356201,
+      0.8606487512588501,
+      0.8730352520942688,
+      0.8025760650634766,
+      0.7727009057998657,
+      0.7014546394348145,
+      0.6866945624351501,
+      0.681542158126831,
+      0.7381680607795715,
+      0.6557687520980835,
+      0.7275815606117249,
+      0.6469742655754089,
+      0.6168563365936279,
+      0.6632050275802612
     ],
     "zipf_loss": [
-      5.888845443725586,
-      2.624917984008789,
-      1.2111440896987915,
-      0.9331302046775818,
-      0.8777217864990234,
-      0.8627199530601501,
-      0.8535035252571106,
-      0.851077675819397,
-      0.8546203970909119,
-      0.8498228788375854,
-      0.8457091450691223,
-      0.8463656902313232,
-      0.8493338823318481,
-      0.8450204133987427,
-      0.8434810042381287,
-      0.8307042717933655,
-      0.8274389505386353,
-      0.8031144142150879,
-      0.7666779160499573,
-      0.6058088541030884,
-      0.3036097288131714,
-      0.20065835118293762,
-      0.184627965092659,
-      0.1593678891658783,
-      0.14444319903850555,
-      0.13536792993545532,
-      0.13580000400543213,
-      0.13336023688316345,
-      0.13025209307670593,
-      0.13264060020446777
+      4.857377529144287,
+      1.2564382553100586,
+      0.8891969919204712,
+      0.8631672859191895,
+      0.8458954691886902,
+      0.8542776107788086,
+      0.8453795909881592,
+      0.8417490720748901,
+      0.8360385894775391,
+      0.8018473386764526,
+      0.30122095346450806,
+      0.14046579599380493,
+      0.11406679451465607,
+      0.1093493178486824,
+      0.0938112735748291,
+      0.08349927514791489,
+      0.06603756546974182,
+      0.06687641143798828,
+      0.06140096113085747,
+      0.06024559587240219,
+      0.0624767541885376,
+      0.05968953296542168,
+      0.0513107106089592,
+      0.05857530236244202,
+      0.047951824963092804,
+      0.047229547053575516,
+      0.06259104609489441,
+      0.051241062581539154,
+      0.049169525504112244,
+      0.06042611971497536
     ],
     "denoise_loss": [],
     "ortho_loss": [
-      0.44732314348220825,
-      0.1550220400094986,
-      0.08860894292593002,
-      0.06823601573705673,
-      0.05723876133561134,
-      0.05476710945367813,
-      0.050798289477825165,
-      0.04427957907319069,
-      0.04448474571108818,
-      0.04900604858994484,
-      0.055344294756650925,
-      0.057683270424604416,
-      0.06224667280912399,
-      0.0660625472664833,
-      0.0682113990187645,
-      0.07531450688838959,
-      0.07825914025306702,
-      0.07614783942699432,
-      0.08581066131591797,
-      0.0885457992553711,
-      0.0975886732339859,
-      0.10165664553642273,
-      0.10155351459980011,
-      0.1034129410982132,
-      0.10390264540910721,
-      0.10435115545988083,
-      0.10648810118436813,
-      0.10846319049596786,
-      0.11003367602825165,
-      0.11091622710227966
+      0.25015148520469666,
+      0.06454251706600189,
+      0.03540093079209328,
+      0.02958545833826065,
+      0.02504236064851284,
+      0.02280796691775322,
+      0.027570093050599098,
+      0.031304821372032166,
+      0.03224661201238632,
+      0.04218565672636032,
+      0.05549744889140129,
+      0.06068537011742592,
+      0.062182605266571045,
+      0.06644897162914276,
+      0.07051082700490952,
+      0.06743474304676056,
+      0.06548100709915161,
+      0.06360627710819244,
+      0.06534555554389954,
+      0.06200597807765007,
+      0.06266725808382034,
+      0.06213154271245003,
+      0.06070336326956749,
+      0.0609060600399971,
+      0.06040602922439575,
+      0.059318866580724716,
+      0.06125802546739578,
+      0.06153956055641174,
+      0.06167338043451309,
+      0.06198073923587799
     ],
     "lr": [
-      1.9600000000000002e-05,
-      3.96e-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,
-      3.944897959183673e-05,
-      3.638775510204082e-05,
-      3.3326530612244897e-05,
-      2.983673469387755e-05,
-      2.6775510204081637e-05,
-      2.3714285714285713e-05,
-      2.022448979591837e-05,
-      1.7163265306122454e-05,
-      1.4102040816326535e-05,
-      1.0612244897959182e-05,
-      7.551020408163262e-06,
-      4.48979591836735e-06
+      3.9200000000000004e-05,
+      7.92e-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,
+      7.889795918367346e-05,
+      7.277551020408164e-05,
+      6.665306122448979e-05,
+      5.96734693877551e-05,
+      5.3551020408163274e-05,
+      4.7428571428571427e-05,
+      4.044897959183674e-05,
+      3.432653061224491e-05,
+      2.820408163265307e-05,
+      2.1224489795918364e-05,
+      1.5102040816326524e-05,
+      8.9795918367347e-06
     ],
     "emb_lr": [],
     "eval_step": [
@@ -280,10 +280,10 @@
       0.0,
       0.02,
       0.01,
-      0.02
+      0.01
     ]
   },
-  "final_accuracy": 0.012083333333333333,
+  "final_accuracy": 0.004583333333333333,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
@@ -298,11 +298,11 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.22148760330578512,
+            "accuracy": 0.2528925619834711,
             "count": 605
           },
           "SS": {
-            "accuracy": 0.6736842105263158,
+            "accuracy": 0.8315789473684211,
             "count": 95
           }
         }
@@ -312,19 +312,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.27450980392156865,
+            "accuracy": 0.30392156862745096,
             "count": 204
           },
           "SC": {
-            "accuracy": 0.1242603550295858,
+            "accuracy": 0.21893491124260356,
             "count": 169
           },
           "SS": {
-            "accuracy": 0.6129032258064516,
+            "accuracy": 0.7096774193548387,
             "count": 31
           },
           "UC": {
-            "accuracy": 0.2533783783783784,
+            "accuracy": 0.27702702702702703,
             "count": 296
           }
         }
@@ -334,23 +334,23 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.3803680981595092,
+            "accuracy": 0.3312883435582822,
             "count": 163
           },
           "SC": {
-            "accuracy": 0.046153846153846156,
+            "accuracy": 0.13846153846153847,
             "count": 130
           },
           "SS": {
-            "accuracy": 0.632183908045977,
+            "accuracy": 0.4482758620689655,
             "count": 87
           },
           "UC": {
-            "accuracy": 0.3448275862068966,
+            "accuracy": 0.3842364532019704,
             "count": 203
           },
           "US": {
-            "accuracy": 0.3247863247863248,
+            "accuracy": 0.5299145299145299,
             "count": 117
           }
         }
@@ -360,23 +360,23 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.4297520661157025,
+            "accuracy": 0.36363636363636365,
             "count": 121
           },
           "SC": {
-            "accuracy": 0.05785123966942149,
+            "accuracy": 0.06611570247933884,
             "count": 121
           },
           "SS": {
-            "accuracy": 0.5714285714285714,
+            "accuracy": 0.42857142857142855,
             "count": 49
           },
           "UC": {
-            "accuracy": 0.3870967741935484,
+            "accuracy": 0.4032258064516129,
             "count": 186
           },
           "US": {
-            "accuracy": 0.4349775784753363,
+            "accuracy": 0.5919282511210763,
             "count": 223
           }
         }
@@ -386,63 +386,63 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.46153846153846156,
+            "accuracy": 0.38461538461538464,
             "count": 104
           },
           "SC": {
-            "accuracy": 0.05660377358490566,
+            "accuracy": 0.11320754716981132,
             "count": 106
           },
           "SS": {
-            "accuracy": 0.6956521739130435,
+            "accuracy": 0.4782608695652174,
             "count": 23
           },
           "UC": {
-            "accuracy": 0.4625,
+            "accuracy": 0.41875,
             "count": 160
           },
           "US": {
-            "accuracy": 0.4169381107491857,
+            "accuracy": 0.49185667752442996,
             "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.0,
+        "full_accuracy": 0.03,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.43,
+            "accuracy": 0.47,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.02,
+            "accuracy": 0.09,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.33,
+            "accuracy": 0.43,
             "count": 100
           },
           "US": {
-            "accuracy": 0.245,
+            "accuracy": 0.3375,
             "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.0,
+        "full_accuracy": 0.12,
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.01,
+            "accuracy": 0.12,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.62,
+            "accuracy": 0.64,
             "count": 100
           },
           "US": {
-            "accuracy": 0.472,
+            "accuracy": 0.636,
             "count": 500
           }
         }
@@ -452,23 +452,23 @@
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.2751677852348993,
+            "accuracy": 0.3087248322147651,
             "count": 447
           },
           "SC": {
-            "accuracy": 0.109375,
+            "accuracy": 0.165625,
             "count": 320
           },
           "SS": {
-            "accuracy": 0.5714285714285714,
+            "accuracy": 0.5892857142857143,
             "count": 56
           },
           "UC": {
-            "accuracy": 0.2684310018903592,
+            "accuracy": 0.29300567107750475,
             "count": 529
           },
           "US": {
-            "accuracy": 0.2708333333333333,
+            "accuracy": 0.4791666666666667,
             "count": 48
           }
         }
@@ -478,19 +478,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.25,
+            "accuracy": 0.35,
             "count": 300
           },
           "SC": {
-            "accuracy": 0.08,
+            "accuracy": 0.1,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.17616580310880828,
+            "accuracy": 0.21243523316062177,
             "count": 193
           },
           "US": {
-            "accuracy": 0.19626168224299065,
+            "accuracy": 0.35514018691588783,
             "count": 107
           }
         }
@@ -500,41 +500,41 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.36,
+            "accuracy": 0.46,
             "count": 200
           },
           "SC": {
-            "accuracy": 0.01,
+            "accuracy": 0.07,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.09765625,
+            "accuracy": 0.15234375,
             "count": 256
           },
           "US": {
-            "accuracy": 0.2013888888888889,
+            "accuracy": 0.2916666666666667,
             "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.0,
+        "full_accuracy": 0.02,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.47,
+            "accuracy": 0.57,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.03,
+            "accuracy": 0.15,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.16993464052287582,
+            "accuracy": 0.2647058823529412,
             "count": 306
           },
           "US": {
-            "accuracy": 0.29896907216494845,
+            "accuracy": 0.4793814432989691,
             "count": 194
           }
         }
@@ -544,15 +544,15 @@
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.03,
+            "accuracy": 0.16,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.30327868852459017,
+            "accuracy": 0.28688524590163933,
             "count": 366
           },
           "US": {
-            "accuracy": 0.5256410256410257,
+            "accuracy": 0.6794871794871795,
             "count": 234
           }
         }
@@ -562,11 +562,11 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.22462562396006655,
+            "accuracy": 0.2579034941763727,
             "count": 601
           },
           "ME": {
-            "accuracy": 0.7171717171717171,
+            "accuracy": 0.898989898989899,
             "count": 99
           }
         }
@@ -576,19 +576,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.4050179211469534,
+            "accuracy": 0.44086021505376344,
             "count": 279
           },
           "MB": {
-            "accuracy": 0.034482758620689655,
+            "accuracy": 0.013793103448275862,
             "count": 145
           },
           "ME": {
-            "accuracy": 0.75,
+            "accuracy": 0.875,
             "count": 24
           },
           "UB": {
-            "accuracy": 0.0992063492063492,
+            "accuracy": 0.15873015873015872,
             "count": 252
           }
         }
@@ -598,23 +598,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6103286384976526,
+            "accuracy": 0.6525821596244131,
             "count": 213
           },
           "MB": {
-            "accuracy": 0.008849557522123894,
+            "accuracy": 0.02654867256637168,
             "count": 113
           },
           "ME": {
-            "accuracy": 0.8352941176470589,
+            "accuracy": 0.8941176470588236,
             "count": 85
           },
           "UB": {
-            "accuracy": 0.15469613259668508,
+            "accuracy": 0.19889502762430938,
             "count": 181
           },
           "UD": {
-            "accuracy": 0.07407407407407407,
+            "accuracy": 0.1111111111111111,
             "count": 108
           }
         }
@@ -624,23 +624,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.7206703910614525,
+            "accuracy": 0.7821229050279329,
             "count": 179
           },
           "MB": {
-            "accuracy": 0.009708737864077669,
+            "accuracy": 0.02912621359223301,
             "count": 103
           },
           "ME": {
-            "accuracy": 0.8571428571428571,
+            "accuracy": 0.8392857142857143,
             "count": 56
           },
           "UB": {
-            "accuracy": 0.15436241610738255,
+            "accuracy": 0.2550335570469799,
             "count": 149
           },
           "UD": {
-            "accuracy": 0.028169014084507043,
+            "accuracy": 0.1267605633802817,
             "count": 213
           }
         }
@@ -650,25 +650,25 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.525,
+            "accuracy": 0.59,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.04,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.3,
+            "accuracy": 0.35,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.07,
+            "accuracy": 0.14333333333333334,
             "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.0,
+        "full_accuracy": 0.02,
         "n_examples": 100,
         "per_subtask": {
           "MD": {
@@ -676,15 +676,15 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.07,
+            "accuracy": 0.09,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.32,
+            "accuracy": 0.52,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.0675,
+            "accuracy": 0.17,
             "count": 400
           }
         }
@@ -694,23 +694,23 @@
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.37833333333333335,
+            "accuracy": 0.4116666666666667,
             "count": 600
           },
           "MB": {
-            "accuracy": 0.018726591760299626,
+            "accuracy": 0.02247191011235955,
             "count": 267
           },
           "ME": {
-            "accuracy": 0.6981132075471698,
+            "accuracy": 0.8679245283018868,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.11845102505694761,
+            "accuracy": 0.1867881548974943,
             "count": 439
           },
           "UD": {
-            "accuracy": 0.04878048780487805,
+            "accuracy": 0.1951219512195122,
             "count": 41
           }
         }
@@ -720,19 +720,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.36,
+            "accuracy": 0.38333333333333336,
             "count": 300
           },
           "MB": {
-            "accuracy": 0.08,
+            "accuracy": 0.01,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.14720812182741116,
+            "accuracy": 0.2131979695431472,
             "count": 197
           },
           "UD": {
-            "accuracy": 0.06796116504854369,
+            "accuracy": 0.038834951456310676,
             "count": 103
           }
         }
@@ -742,7 +742,7 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.51,
+            "accuracy": 0.555,
             "count": 200
           },
           "MB": {
@@ -750,11 +750,11 @@
             "count": 100
           },
           "UB": {
-            "accuracy": 0.15384615384615385,
+            "accuracy": 0.21862348178137653,
             "count": 247
           },
           "UD": {
-            "accuracy": 0.06535947712418301,
+            "accuracy": 0.0784313725490196,
             "count": 153
           }
         }
@@ -768,22 +768,22 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.05,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.10738255033557047,
+            "accuracy": 0.174496644295302,
             "count": 298
           },
           "UD": {
-            "accuracy": 0.15346534653465346,
+            "accuracy": 0.04455445544554455,
             "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.0,
+      "overall_accuracy": 0.007916666666666667,
       "total_examples": 2400,
       "n_splits": 22
     }
@@ -798,15 +798,15 @@
     },
     "splits": {
       "add_S0": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.5553719008264463,
+            "accuracy": 0.39669421487603307,
             "count": 605
           },
           "SS": {
-            "accuracy": 0.9052631578947369,
+            "accuracy": 0.9263157894736842,
             "count": 95
           }
         }
@@ -816,163 +816,163 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.5294117647058824,
+            "accuracy": 0.45098039215686275,
             "count": 204
           },
           "SC": {
-            "accuracy": 0.33727810650887574,
+            "accuracy": 0.2603550295857988,
             "count": 169
           },
           "SS": {
-            "accuracy": 0.7419354838709677,
+            "accuracy": 0.8709677419354839,
             "count": 31
           },
           "UC": {
-            "accuracy": 0.44594594594594594,
+            "accuracy": 0.3141891891891892,
             "count": 296
           }
         }
       },
       "add_S2": {
-        "full_accuracy": 0.01,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.6441717791411042,
+            "accuracy": 0.4785276073619632,
             "count": 163
           },
           "SC": {
-            "accuracy": 0.2923076923076923,
+            "accuracy": 0.12307692307692308,
             "count": 130
           },
           "SS": {
-            "accuracy": 0.7701149425287356,
+            "accuracy": 0.3563218390804598,
             "count": 87
           },
           "UC": {
-            "accuracy": 0.4729064039408867,
+            "accuracy": 0.4187192118226601,
             "count": 203
           },
           "US": {
-            "accuracy": 0.5128205128205128,
+            "accuracy": 0.5641025641025641,
             "count": 117
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 0.03,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.6694214876033058,
+            "accuracy": 0.512396694214876,
             "count": 121
           },
           "SC": {
-            "accuracy": 0.256198347107438,
+            "accuracy": 0.09917355371900827,
             "count": 121
           },
           "SS": {
-            "accuracy": 0.8367346938775511,
+            "accuracy": 0.5102040816326531,
             "count": 49
           },
           "UC": {
-            "accuracy": 0.5268817204301075,
+            "accuracy": 0.4032258064516129,
             "count": 186
           },
           "US": {
-            "accuracy": 0.45739910313901344,
+            "accuracy": 0.5560538116591929,
             "count": 223
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 0.04,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.7019230769230769,
+            "accuracy": 0.5865384615384616,
             "count": 104
           },
           "SC": {
-            "accuracy": 0.36792452830188677,
+            "accuracy": 0.11320754716981132,
             "count": 106
           },
           "SS": {
-            "accuracy": 0.9130434782608695,
+            "accuracy": 0.6521739130434783,
             "count": 23
           },
           "UC": {
-            "accuracy": 0.525,
+            "accuracy": 0.43125,
             "count": 160
           },
           "US": {
-            "accuracy": 0.40716612377850164,
+            "accuracy": 0.48534201954397393,
             "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.05,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.67,
+            "accuracy": 0.63,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.37,
+            "accuracy": 0.06,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.38,
+            "accuracy": 0.34,
             "count": 100
           },
           "US": {
-            "accuracy": 0.27,
+            "accuracy": 0.2575,
             "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.11,
+        "full_accuracy": 0.1,
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.23,
+            "accuracy": 0.1,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.39,
+            "accuracy": 0.53,
             "count": 100
           },
           "US": {
-            "accuracy": 0.32,
+            "accuracy": 0.506,
             "count": 500
           }
         }
       },
       "add_random": {
-        "full_accuracy": 0.005,
+        "full_accuracy": 0.0,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.5391498881431768,
+            "accuracy": 0.46308724832214765,
             "count": 447
           },
           "SC": {
-            "accuracy": 0.328125,
+            "accuracy": 0.1875,
             "count": 320
           },
           "SS": {
-            "accuracy": 0.75,
+            "accuracy": 0.6785714285714286,
             "count": 56
           },
           "UC": {
-            "accuracy": 0.4612476370510397,
+            "accuracy": 0.3856332703213611,
             "count": 529
           },
           "US": {
-            "accuracy": 0.3541666666666667,
+            "accuracy": 0.4791666666666667,
             "count": 48
           }
         }
@@ -982,19 +982,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.67,
+            "accuracy": 0.47333333333333333,
             "count": 300
           },
           "SC": {
-            "accuracy": 0.22,
+            "accuracy": 0.04,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.3626943005181347,
+            "accuracy": 0.21761658031088082,
             "count": 193
           },
           "US": {
-            "accuracy": 0.29906542056074764,
+            "accuracy": 0.2336448598130841,
             "count": 107
           }
         }
@@ -1004,19 +1004,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.755,
+            "accuracy": 0.625,
             "count": 200
           },
           "SC": {
-            "accuracy": 0.32,
+            "accuracy": 0.04,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.33984375,
+            "accuracy": 0.19140625,
             "count": 256
           },
           "US": {
-            "accuracy": 0.3125,
+            "accuracy": 0.2986111111111111,
             "count": 144
           }
         }
@@ -1026,19 +1026,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.67,
+            "accuracy": 0.62,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.42,
+            "accuracy": 0.08,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.45098039215686275,
+            "accuracy": 0.25163398692810457,
             "count": 306
           },
           "US": {
-            "accuracy": 0.37628865979381443,
+            "accuracy": 0.38144329896907214,
             "count": 194
           }
         }
@@ -1048,15 +1048,15 @@
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.31,
+            "accuracy": 0.14,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.5081967213114754,
+            "accuracy": 0.32786885245901637,
             "count": 366
           },
           "US": {
-            "accuracy": 0.6495726495726496,
+            "accuracy": 0.7307692307692307,
             "count": 234
           }
         }
@@ -1066,11 +1066,11 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.540765391014975,
+            "accuracy": 0.36605657237936773,
             "count": 601
           },
           "ME": {
-            "accuracy": 0.7575757575757576,
+            "accuracy": 0.8888888888888888,
             "count": 99
           }
         }
@@ -1080,19 +1080,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6523297491039427,
+            "accuracy": 0.5913978494623656,
             "count": 279
           },
           "MB": {
-            "accuracy": 0.10344827586206896,
+            "accuracy": 0.034482758620689655,
             "count": 145
           },
           "ME": {
-            "accuracy": 0.7916666666666666,
+            "accuracy": 0.75,
             "count": 24
           },
           "UB": {
-            "accuracy": 0.3333333333333333,
+            "accuracy": 0.2222222222222222,
             "count": 252
           }
         }
@@ -1102,23 +1102,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.8169014084507042,
+            "accuracy": 0.704225352112676,
             "count": 213
           },
           "MB": {
-            "accuracy": 0.12389380530973451,
+            "accuracy": 0.04424778761061947,
             "count": 113
           },
           "ME": {
-            "accuracy": 0.9529411764705882,
+            "accuracy": 0.9647058823529412,
             "count": 85
           },
           "UB": {
-            "accuracy": 0.27071823204419887,
+            "accuracy": 0.281767955801105,
             "count": 181
           },
           "UD": {
-            "accuracy": 0.1111111111111111,
+            "accuracy": 0.12037037037037036,
             "count": 108
           }
         }
@@ -1128,23 +1128,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.8659217877094972,
+            "accuracy": 0.8156424581005587,
             "count": 179
           },
           "MB": {
-            "accuracy": 0.10679611650485436,
+            "accuracy": 0.0,
             "count": 103
           },
           "ME": {
-            "accuracy": 0.9642857142857143,
+            "accuracy": 0.9285714285714286,
             "count": 56
           },
           "UB": {
-            "accuracy": 0.30201342281879195,
+            "accuracy": 0.348993288590604,
             "count": 149
           },
           "UD": {
-            "accuracy": 0.09859154929577464,
+            "accuracy": 0.0892018779342723,
             "count": 213
           }
         }
@@ -1154,25 +1154,25 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.79,
+            "accuracy": 0.76,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.15,
+            "accuracy": 0.01,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.4,
+            "accuracy": 0.44,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.11,
+            "accuracy": 0.08333333333333333,
             "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.01,
         "n_examples": 100,
         "per_subtask": {
           "MD": {
@@ -1184,11 +1184,11 @@
             "count": 100
           },
           "UB": {
-            "accuracy": 0.56,
+            "accuracy": 0.7,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.095,
+            "accuracy": 0.065,
             "count": 400
           }
         }
@@ -1198,23 +1198,23 @@
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6883333333333334,
+            "accuracy": 0.555,
             "count": 600
           },
           "MB": {
-            "accuracy": 0.12359550561797752,
+            "accuracy": 0.03745318352059925,
             "count": 267
           },
           "ME": {
-            "accuracy": 0.7924528301886793,
+            "accuracy": 0.9245283018867925,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.3120728929384966,
+            "accuracy": 0.275626423690205,
             "count": 439
           },
           "UD": {
-            "accuracy": 0.17073170731707318,
+            "accuracy": 0.1951219512195122,
             "count": 41
           }
         }
@@ -1224,19 +1224,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6766666666666666,
+            "accuracy": 0.55,
             "count": 300
           },
           "MB": {
-            "accuracy": 0.12,
+            "accuracy": 0.07,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.24873096446700507,
+            "accuracy": 0.2233502538071066,
             "count": 197
           },
           "UD": {
-            "accuracy": 0.11650485436893204,
+            "accuracy": 0.1941747572815534,
             "count": 103
           }
         }
@@ -1246,19 +1246,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.775,
+            "accuracy": 0.68,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.18,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.2874493927125506,
+            "accuracy": 0.2550607287449393,
             "count": 247
           },
           "UD": {
-            "accuracy": 0.1895424836601307,
+            "accuracy": 0.1568627450980392,
             "count": 153
           }
         }
@@ -1272,26 +1272,26 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.08,
+            "accuracy": 0.06,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.30201342281879195,
+            "accuracy": 0.31208053691275167,
             "count": 298
           },
           "UD": {
-            "accuracy": 0.06930693069306931,
+            "accuracy": 0.19801980198019803,
             "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.012083333333333333,
+      "overall_accuracy": 0.004583333333333333,
       "total_examples": 2400,
       "n_splits": 22
     }
   },
-  "sorl_overall_accuracy": 0.012083333333333333,
-  "sft_overall_accuracy": 0.0
+  "sorl_overall_accuracy": 0.004583333333333333,
+  "sft_overall_accuracy": 0.007916666666666667
 }

add_sub_sorl_v1_abs30_10K/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f6e5e1e62f7dc6b60e3669e47c1843dcad77c3bdb610639bf38311f70d6d063
+oid sha256:688cc5ac307149019958195a1925a9ceaf23aacc9763b9e0e85256c740254636
 size 650385300

add_sub_sorl_v1_abs30_10K/train_config.json

@@ -17,7 +17,7 @@
   "target_vocab_util": 0.8,
   "min_abs_ppl": 0.0,
   "zipf_alpha": 1.0,
-  "lr": 4e-05,
+  "lr": 8e-05,
   "emb_lr_mult": 1.0,
   "weight_decay": 0.01,
   "warmup_steps": 100,
@@ -36,7 +36,7 @@
   "eval_every": 156,
   "save_every": 999999,
   "eval_samples": 100,
-  "output_dir": "ckpt/sweep/as_sorl_abs30_K4_10K",
+  "output_dir": "ckpt/sweep/as_sorl_abs30_10K",
   "eval_K": 4,
   "alpha_traj": 0.0,
   "corrupt_method": "shuffle",
@@ -69,16 +69,16 @@
   "no_wandb": false,
   "n_params": 162519662,
   "run_name": "add_sub_sorl_v1_abs30_10K",
-  "git_commit": "f447da529caceac8c7d256cbb2cd185cbc50feac",
-  "timestamp": "2026-04-12T09:21:17.547661+00:00",
+  "git_commit": "dc8dd776fb0c30a4c9073052dcc5e943e0fd80c6",
+  "timestamp": "2026-04-13T07:34:34.157977+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": "owo1a52f",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/owo1a52f",
-  "final_accuracy": 0.012083333333333333,
-  "sft_accuracy": 0.0,
+  "wandb_run_id": "4pjkd149",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/4pjkd149",
+  "final_accuracy": 0.004583333333333333,
+  "sft_accuracy": 0.007916666666666667,
   "eval_method": "ArithmeticEvaluator"
 }