Upload add_sub_sorl_v1_abs10_10K

add_sub_sorl_v1_abs10_10K/metrics.json

@@ -33,229 +33,229 @@
       1563
     ],
     "loss": [
-      13.724945068359375,
-      7.864343166351318,
-      5.013725280761719,
-      2.741554021835327,
-      2.641360282897949,
-      2.458710193634033,
-      2.5750374794006348,
-      2.7018885612487793,
-      2.5789809226989746,
-      2.3690404891967773,
-      1.7872730493545532,
-      1.8722014427185059,
-      1.4984050989151,
-      1.2242059707641602,
-      1.2751460075378418,
-      1.0107839107513428,
-      1.6164339780807495,
-      0.9860085248947144,
-      0.9987770915031433,
-      1.071738600730896,
-      0.49998074769973755,
-      0.004859551787376404,
-      -0.1695592701435089,
-      0.0747871994972229,
-      -0.27059921622276306,
-      -0.13733237981796265,
-      0.056721221655607224,
-      0.051335956901311874,
-      -0.45392414927482605,
-      -0.3348205089569092
+      11.961856842041016,
+      5.24287223815918,
+      3.2151432037353516,
+      2.6320042610168457,
+      2.6821465492248535,
+      2.0658371448516846,
+      1.942716360092163,
+      1.8816739320755005,
+      1.884223461151123,
+      0.5458273887634277,
+      -0.02370402216911316,
+      -0.10736316442489624,
+      -0.0433916300535202,
+      -0.08627352118492126,
+      0.19944323599338531,
+      -0.05154527723789215,
+      0.18322159349918365,
+      0.12385116517543793,
+      0.17138411104679108,
+      -0.045592449605464935,
+      0.21150481700897217,
+      0.0073037222027778625,
+      -0.3383682370185852,
+      0.12239806354045868,
+      -0.13985984027385712,
+      0.046681616455316544,
+      0.25705066323280334,
+      0.1217120885848999,
+      0.1391371488571167,
+      -0.13681045174598694
     ],
     "base_loss": [
-      7.9022040367126465,
-      5.812614917755127,
-      3.858829975128174,
-      2.114943742752075,
-      1.96530020236969,
-      1.9036047458648682,
-      1.893314242362976,
-      1.8308947086334229,
-      1.8680362701416016,
-      1.7989755868911743,
-      1.8180350065231323,
-      1.8677423000335693,
-      1.862571120262146,
-      1.8560221195220947,
-      1.838614583015442,
-      1.7637040615081787,
-      1.8746064901351929,
-      1.8687989711761475,
-      1.8472496271133423,
-      1.8745392560958862,
-      1.8059179782867432,
-      1.755523920059204,
-      1.7115023136138916,
-      1.7793734073638916,
-      1.759653091430664,
-      1.7252968549728394,
-      1.7201368808746338,
-      1.7096773386001587,
-      1.7525498867034912,
-      1.710257887840271
+      6.969983100891113,
+      3.7504212856292725,
+      1.9797016382217407,
+      1.8771779537200928,
+      1.8847401142120361,
+      1.8795526027679443,
+      1.8393981456756592,
+      1.8044995069503784,
+      1.8537063598632812,
+      1.7935826778411865,
+      1.8110519647598267,
+      1.8323622941970825,
+      1.8425616025924683,
+      1.8131767511367798,
+      1.8040305376052856,
+      1.721121907234192,
+      1.8190782070159912,
+      1.812418818473816,
+      1.7917753458023071,
+      1.793439507484436,
+      1.725421667098999,
+      1.7205660343170166,
+      1.6890180110931396,
+      1.746482014656067,
+      1.7506285905838013,
+      1.6954907178878784,
+      1.7080529928207397,
+      1.6827865839004517,
+      1.688535451889038,
+      1.6724849939346313
     ],
     "info_loss": [
-      -0.2345585823059082,
-      -0.18186569213867188,
-      -0.11677074432373047,
-      -0.13591337203979492,
-      -0.12461328506469727,
-      -0.13464367389678955,
-      -0.12099599838256836,
-      -0.10128498077392578,
-      -0.11677730083465576,
-      -0.13061606884002686,
-      -0.1908276081085205,
-      -0.18726766109466553,
-      -0.2237793207168579,
-      -0.25037193298339844,
-      -0.24307358264923096,
-      -0.26141083240509033,
-      -0.2112140655517578,
-      -0.2714040279388428,
-      -0.2562897205352783,
-      -0.21821486949920654,
-      -0.1999720335006714,
-      -0.19893121719360352,
-      -0.20633161067962646,
-      -0.1859527826309204,
-      -0.21592140197753906,
-      -0.19767189025878906,
-      -0.17773842811584473,
-      -0.17674016952514648,
-      -0.2308502197265625,
-      -0.21450459957122803
+      -0.18472051620483398,
+      -0.08357405662536621,
+      -0.06814658641815186,
+      -0.11294722557067871,
+      -0.10811793804168701,
+      -0.16878092288970947,
+      -0.17634761333465576,
+      -0.17674624919891357,
+      -0.16523611545562744,
+      -0.19186913967132568,
+      -0.2089073657989502,
+      -0.20613396167755127,
+      -0.1997373104095459,
+      -0.19908201694488525,
+      -0.16838586330413818,
+      -0.1851973533630371,
+      -0.1715489625930786,
+      -0.17607414722442627,
+      -0.16980135440826416,
+      -0.1923583745956421,
+      -0.15899121761322021,
+      -0.1772533655166626,
+      -0.20877563953399658,
+      -0.16812872886657715,
+      -0.19421494007110596,
+      -0.1697760820388794,
+      -0.1513657569885254,
+      -0.16188645362854004,
+      -0.15964651107788086,
+      -0.18581724166870117
     ],
     "abs_loss": [
-      2.2656421661376953,
-      2.060791254043579,
-      1.8972232341766357,
-      1.8288743495941162,
-      1.8518760204315186,
-      1.8478550910949707,
-      1.8548732995986938,
-      1.8174316883087158,
-      1.8104970455169678,
-      1.8124628067016602,
-      1.8303263187408447,
-      1.860059380531311,
-      1.8033771514892578,
-      1.8484278917312622,
-      1.7977077960968018,
-      1.7986536026000977,
-      1.8080449104309082,
-      1.7756034135818481,
-      1.678572177886963,
-      1.2864962816238403,
-      0.798673152923584,
-      0.611975371837616,
-      0.5902649164199829,
-      0.584165632724762,
-      0.5656200647354126,
-      0.5475164651870728,
-      0.5673137903213501,
-      0.5301952362060547,
-      0.4829615652561188,
-      0.5067222118377686
+      2.1909255981445312,
+      1.8499255180358887,
+      1.8133636713027954,
+      1.8108993768692017,
+      1.8413399457931519,
+      1.8359529972076416,
+      1.8084602355957031,
+      1.7734503746032715,
+      1.6312320232391357,
+      1.0209534168243408,
+      0.7137269377708435,
+      0.5659414529800415,
+      0.4661961793899536,
+      0.4323907196521759,
+      0.36104658246040344,
+      0.4403992295265198,
+      0.4302736222743988,
+      0.42298078536987305,
+      0.34669893980026245,
+      0.45547735691070557,
+      0.4821462035179138,
+      0.40443599224090576,
+      0.44047102332115173,
+      0.37605512142181396,
+      0.35405462980270386,
+      0.33734890818595886,
+      0.3338230848312378,
+      0.30671998858451843,
+      0.26178568601608276,
+      0.33180034160614014
     ],
     "zipf_loss": [
-      7.941762447357178,
-      3.664306163787842,
-      2.132880449295044,
-      1.8028565645217896,
-      1.7370052337646484,
-      1.7167567014694214,
-      1.7061958312988281,
-      1.7021006345748901,
-      1.6976680755615234,
-      1.694979190826416,
-      1.6944814920425415,
-      1.6911298036575317,
-      1.6932895183563232,
-      1.6870603561401367,
-      1.6874964237213135,
-      1.6813228130340576,
-      1.673163652420044,
-      1.6536895036697388,
-      1.546567440032959,
-      1.250698447227478,
-      0.6139158010482788,
-      0.1774502694606781,
-      0.1232280433177948,
-      0.09652505815029144,
-      0.07239970564842224,
-      0.059338025748729706,
-      0.05723724141716957,
-      0.05604078993201256,
-      0.053732018917798996,
-      0.04929535463452339
+      6.619986534118652,
+      2.1431989669799805,
+      1.7355711460113525,
+      1.7032086849212646,
+      1.6944518089294434,
+      1.6904984712600708,
+      1.685948371887207,
+      1.6672918796539307,
+      1.5197550058364868,
+      0.568840742111206,
+      0.18294498324394226,
+      0.06502000987529755,
+      0.06480024755001068,
+      0.0481308251619339,
+      0.04316667094826698,
+      0.03526642918586731,
+      0.03660564124584198,
+      0.0298757404088974,
+      0.042952414602041245,
+      0.03900405019521713,
+      0.027780719101428986,
+      0.01882774382829666,
+      0.016323039308190346,
+      0.01959783211350441,
+      0.016255516558885574,
+      0.015216827392578125,
+      0.029272940009832382,
+      0.02711804211139679,
+      0.020888233557343483,
+      0.0156969353556633
     ],
     "denoise_loss": [],
     "ortho_loss": [
-      0.5284707546234131,
-      0.21797625720500946,
-      0.12806634604930878,
-      0.09031304717063904,
-      0.08532247692346573,
-      0.07537004351615906,
-      0.07645349949598312,
-      0.07695373892784119,
-      0.07142108678817749,
-      0.08406423032283783,
-      0.08843859285116196,
-      0.07815887778997421,
-      0.07982385903596878,
-      0.07472201436758041,
-      0.07963292300701141,
-      0.07144557684659958,
-      0.07161661982536316,
-      0.07070355117321014,
-      0.06902579218149185,
-      0.06851960718631744,
-      0.06775054335594177,
-      0.0655888020992279,
-      0.059131793677806854,
-      0.05900496616959572,
-      0.056693185120821,
-      0.05708722025156021,
-      0.058303531259298325,
-      0.059081800282001495,
-      0.059238508343696594,
-      0.059114448726177216
+      0.3513818085193634,
+      0.12336498498916626,
+      0.07400115579366684,
+      0.04318969324231148,
+      0.03977782279253006,
+      0.031715553253889084,
+      0.03395743668079376,
+      0.037701401859521866,
+      0.03583124652504921,
+      0.0376955084502697,
+      0.03476428985595703,
+      0.03290867805480957,
+      0.03230033069849014,
+      0.028476815670728683,
+      0.029924070462584496,
+      0.025573937222361565,
+      0.023195529356598854,
+      0.024144373834133148,
+      0.026353251188993454,
+      0.02235715091228485,
+      0.025194751098752022,
+      0.025079840794205666,
+      0.028410209342837334,
+      0.03212076798081398,
+      0.032003507018089294,
+      0.03370780497789383,
+      0.03460211679339409,
+      0.03515707328915596,
+      0.03469262272119522,
+      0.03475438058376312
     ],
     "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,
-      0.01
+      0.0
     ]
   },
-  "final_accuracy": 0.008333333333333333,
+  "final_accuracy": 0.007083333333333333,
   "sft_eval": {
     "config": {
       "ops": "add_sub",
@@ -298,11 +298,11 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.24297520661157024,
+            "accuracy": 0.30247933884297523,
             "count": 605
           },
           "SS": {
-            "accuracy": 0.8736842105263158,
+            "accuracy": 0.8,
             "count": 95
           }
         }
@@ -312,19 +312,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.28921568627450983,
+            "accuracy": 0.29411764705882354,
             "count": 204
           },
           "SC": {
-            "accuracy": 0.14792899408284024,
+            "accuracy": 0.20118343195266272,
             "count": 169
           },
           "SS": {
-            "accuracy": 0.8064516129032258,
+            "accuracy": 0.5483870967741935,
             "count": 31
           },
           "UC": {
-            "accuracy": 0.23648648648648649,
+            "accuracy": 0.2905405405405405,
             "count": 296
           }
         }
@@ -334,23 +334,23 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.3619631901840491,
+            "accuracy": 0.4110429447852761,
             "count": 163
           },
           "SC": {
-            "accuracy": 0.1,
+            "accuracy": 0.14615384615384616,
             "count": 130
           },
           "SS": {
-            "accuracy": 0.4827586206896552,
+            "accuracy": 0.41379310344827586,
             "count": 87
           },
           "UC": {
-            "accuracy": 0.33497536945812806,
+            "accuracy": 0.4187192118226601,
             "count": 203
           },
           "US": {
-            "accuracy": 0.5384615384615384,
+            "accuracy": 0.5726495726495726,
             "count": 117
           }
         }
@@ -360,23 +360,23 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.4132231404958678,
+            "accuracy": 0.4214876033057851,
             "count": 121
           },
           "SC": {
-            "accuracy": 0.06611570247933884,
+            "accuracy": 0.09090909090909091,
             "count": 121
           },
           "SS": {
-            "accuracy": 0.46938775510204084,
+            "accuracy": 0.3673469387755102,
             "count": 49
           },
           "UC": {
-            "accuracy": 0.34946236559139787,
+            "accuracy": 0.41935483870967744,
             "count": 186
           },
           "US": {
-            "accuracy": 0.5650224215246636,
+            "accuracy": 0.6278026905829597,
             "count": 223
           }
         }
@@ -386,33 +386,33 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.4519230769230769,
+            "accuracy": 0.41346153846153844,
             "count": 104
           },
           "SC": {
-            "accuracy": 0.08490566037735849,
+            "accuracy": 0.1320754716981132,
             "count": 106
           },
           "SS": {
-            "accuracy": 0.6086956521739131,
+            "accuracy": 0.4782608695652174,
             "count": 23
           },
           "UC": {
-            "accuracy": 0.40625,
+            "accuracy": 0.5125,
             "count": 160
           },
           "US": {
-            "accuracy": 0.4560260586319218,
+            "accuracy": 0.6319218241042345,
             "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.01,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.43,
+            "accuracy": 0.49,
             "count": 100
           },
           "SC": {
@@ -420,29 +420,29 @@
             "count": 100
           },
           "UC": {
-            "accuracy": 0.51,
+            "accuracy": 0.67,
             "count": 100
           },
           "US": {
-            "accuracy": 0.36,
+            "accuracy": 0.44,
             "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.05,
+        "full_accuracy": 0.14,
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.05,
+            "accuracy": 0.14,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.53,
+            "accuracy": 0.77,
             "count": 100
           },
           "US": {
-            "accuracy": 0.502,
+            "accuracy": 0.746,
             "count": 500
           }
         }
@@ -452,23 +452,23 @@
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.27293064876957496,
+            "accuracy": 0.3243847874720358,
             "count": 447
           },
           "SC": {
-            "accuracy": 0.1125,
+            "accuracy": 0.14375,
             "count": 320
           },
           "SS": {
-            "accuracy": 0.6964285714285714,
+            "accuracy": 0.6428571428571429,
             "count": 56
           },
           "UC": {
-            "accuracy": 0.2665406427221172,
+            "accuracy": 0.32514177693761814,
             "count": 529
           },
           "US": {
-            "accuracy": 0.5,
+            "accuracy": 0.5208333333333334,
             "count": 48
           }
         }
@@ -478,19 +478,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.31666666666666665,
+            "accuracy": 0.37333333333333335,
             "count": 300
           },
           "SC": {
-            "accuracy": 0.07,
+            "accuracy": 0.1,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.19689119170984457,
+            "accuracy": 0.24870466321243523,
             "count": 193
           },
           "US": {
-            "accuracy": 0.2523364485981308,
+            "accuracy": 0.29906542056074764,
             "count": 107
           }
         }
@@ -500,41 +500,41 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.405,
+            "accuracy": 0.46,
             "count": 200
           },
           "SC": {
-            "accuracy": 0.05,
+            "accuracy": 0.02,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.1484375,
+            "accuracy": 0.203125,
             "count": 256
           },
           "US": {
-            "accuracy": 0.2986111111111111,
+            "accuracy": 0.2916666666666667,
             "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.01,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.54,
+            "accuracy": 0.53,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.1,
+            "accuracy": 0.09,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.24836601307189543,
+            "accuracy": 0.2875816993464052,
             "count": 306
           },
           "US": {
-            "accuracy": 0.5,
+            "accuracy": 0.5670103092783505,
             "count": 194
           }
         }
@@ -548,11 +548,11 @@
             "count": 100
           },
           "UC": {
-            "accuracy": 0.2459016393442623,
+            "accuracy": 0.29508196721311475,
             "count": 366
           },
           "US": {
-            "accuracy": 0.6709401709401709,
+            "accuracy": 0.7948717948717948,
             "count": 234
           }
         }
@@ -562,11 +562,11 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.23960066555740434,
+            "accuracy": 0.22462562396006655,
             "count": 601
           },
           "ME": {
-            "accuracy": 0.8282828282828283,
+            "accuracy": 0.98989898989899,
             "count": 99
           }
         }
@@ -576,19 +576,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.4444444444444444,
+            "accuracy": 0.4050179211469534,
             "count": 279
           },
           "MB": {
-            "accuracy": 0.034482758620689655,
+            "accuracy": 0.013793103448275862,
             "count": 145
           },
           "ME": {
-            "accuracy": 0.875,
+            "accuracy": 0.9583333333333334,
             "count": 24
           },
           "UB": {
-            "accuracy": 0.1388888888888889,
+            "accuracy": 0.11507936507936507,
             "count": 252
           }
         }
@@ -598,23 +598,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6291079812206573,
+            "accuracy": 0.6244131455399061,
             "count": 213
           },
           "MB": {
-            "accuracy": 0.035398230088495575,
+            "accuracy": 0.017699115044247787,
             "count": 113
           },
           "ME": {
-            "accuracy": 0.8705882352941177,
+            "accuracy": 0.9529411764705882,
             "count": 85
           },
           "UB": {
-            "accuracy": 0.19337016574585636,
+            "accuracy": 0.15469613259668508,
             "count": 181
           },
           "UD": {
-            "accuracy": 0.1388888888888889,
+            "accuracy": 0.037037037037037035,
             "count": 108
           }
         }
@@ -624,23 +624,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.7374301675977654,
+            "accuracy": 0.7597765363128491,
             "count": 179
           },
           "MB": {
-            "accuracy": 0.02912621359223301,
+            "accuracy": 0.009708737864077669,
             "count": 103
           },
           "ME": {
-            "accuracy": 0.8928571428571429,
+            "accuracy": 0.9464285714285714,
             "count": 56
           },
           "UB": {
-            "accuracy": 0.2214765100671141,
+            "accuracy": 0.19463087248322147,
             "count": 149
           },
           "UD": {
-            "accuracy": 0.13145539906103287,
+            "accuracy": 0.06572769953051644,
             "count": 213
           }
         }
@@ -650,19 +650,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.55,
+            "accuracy": 0.515,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.06,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.35,
+            "accuracy": 0.33,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.16666666666666666,
+            "accuracy": 0.1,
             "count": 300
           }
         }
@@ -676,15 +676,15 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.04,
+            "accuracy": 0.02,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.47,
+            "accuracy": 0.38,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.16,
+            "accuracy": 0.05,
             "count": 400
           }
         }
@@ -694,23 +694,23 @@
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.4066666666666667,
+            "accuracy": 0.36666666666666664,
             "count": 600
           },
           "MB": {
-            "accuracy": 0.011235955056179775,
+            "accuracy": 0.0149812734082397,
             "count": 267
           },
           "ME": {
-            "accuracy": 0.7547169811320755,
+            "accuracy": 0.9811320754716981,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.1662870159453303,
+            "accuracy": 0.13439635535307518,
             "count": 439
           },
           "UD": {
-            "accuracy": 0.21951219512195122,
+            "accuracy": 0.0975609756097561,
             "count": 41
           }
         }
@@ -720,19 +720,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.38,
+            "accuracy": 0.36,
             "count": 300
           },
           "MB": {
-            "accuracy": 0.09,
+            "accuracy": 0.05,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.16243654822335024,
+            "accuracy": 0.18781725888324874,
             "count": 197
           },
           "UD": {
-            "accuracy": 0.20388349514563106,
+            "accuracy": 0.0970873786407767,
             "count": 103
           }
         }
@@ -742,19 +742,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.55,
+            "accuracy": 0.52,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.07,
+            "accuracy": 0.02,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.17408906882591094,
+            "accuracy": 0.16194331983805668,
             "count": 247
           },
           "UD": {
-            "accuracy": 0.1568627450980392,
+            "accuracy": 0.0784313725490196,
             "count": 153
           }
         }
@@ -768,22 +768,22 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.04,
+            "accuracy": 0.02,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.1644295302013423,
+            "accuracy": 0.13087248322147652,
             "count": 298
           },
           "UD": {
-            "accuracy": 0.1188118811881188,
+            "accuracy": 0.10396039603960396,
             "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.004166666666666667,
+      "overall_accuracy": 0.007083333333333333,
       "total_examples": 2400,
       "n_splits": 22
     }
@@ -802,11 +802,11 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.3256198347107438,
+            "accuracy": 0.33884297520661155,
             "count": 605
           },
           "SS": {
-            "accuracy": 0.968421052631579,
+            "accuracy": 0.7578947368421053,
             "count": 95
           }
         }
@@ -816,19 +816,19 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.3431372549019608,
+            "accuracy": 0.3333333333333333,
             "count": 204
           },
           "SC": {
-            "accuracy": 0.23076923076923078,
+            "accuracy": 0.2781065088757396,
             "count": 169
           },
           "SS": {
-            "accuracy": 0.8387096774193549,
+            "accuracy": 0.6129032258064516,
             "count": 31
           },
           "UC": {
-            "accuracy": 0.32094594594594594,
+            "accuracy": 0.3310810810810811,
             "count": 296
           }
         }
@@ -838,11 +838,11 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.4785276073619632,
+            "accuracy": 0.34355828220858897,
             "count": 163
           },
           "SC": {
-            "accuracy": 0.13076923076923078,
+            "accuracy": 0.15384615384615385,
             "count": 130
           },
           "SS": {
@@ -850,7 +850,7 @@
             "count": 87
           },
           "UC": {
-            "accuracy": 0.4039408866995074,
+            "accuracy": 0.4236453201970443,
             "count": 203
           },
           "US": {
@@ -860,27 +860,27 @@
         }
       },
       "add_S3": {
-        "full_accuracy": 0.0,
+        "full_accuracy": 0.01,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.5289256198347108,
+            "accuracy": 0.4628099173553719,
             "count": 121
           },
           "SC": {
-            "accuracy": 0.0743801652892562,
+            "accuracy": 0.17355371900826447,
             "count": 121
           },
           "SS": {
-            "accuracy": 0.4897959183673469,
+            "accuracy": 0.5306122448979592,
             "count": 49
           },
           "UC": {
-            "accuracy": 0.41935483870967744,
+            "accuracy": 0.42473118279569894,
             "count": 186
           },
           "US": {
-            "accuracy": 0.5964125560538116,
+            "accuracy": 0.726457399103139,
             "count": 223
           }
         }
@@ -890,29 +890,29 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.5480769230769231,
+            "accuracy": 0.40384615384615385,
             "count": 104
           },
           "SC": {
-            "accuracy": 0.08490566037735849,
+            "accuracy": 0.16037735849056603,
             "count": 106
           },
           "SS": {
-            "accuracy": 0.6521739130434783,
+            "accuracy": 0.5652173913043478,
             "count": 23
           },
           "UC": {
-            "accuracy": 0.41875,
+            "accuracy": 0.46875,
             "count": 160
           },
           "US": {
-            "accuracy": 0.5374592833876222,
+            "accuracy": 0.6775244299674267,
             "count": 307
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
@@ -920,33 +920,33 @@
             "count": 100
           },
           "SC": {
-            "accuracy": 0.05,
+            "accuracy": 0.1,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.46,
+            "accuracy": 0.57,
             "count": 100
           },
           "US": {
-            "accuracy": 0.3175,
+            "accuracy": 0.4925,
             "count": 400
           }
         }
       },
       "add_S6": {
-        "full_accuracy": 0.12,
+        "full_accuracy": 0.15,
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.12,
+            "accuracy": 0.15,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.48,
+            "accuracy": 0.7,
             "count": 100
           },
           "US": {
-            "accuracy": 0.448,
+            "accuracy": 0.682,
             "count": 500
           }
         }
@@ -956,23 +956,23 @@
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.36017897091722595,
+            "accuracy": 0.38478747203579416,
             "count": 447
           },
           "SC": {
-            "accuracy": 0.18125,
+            "accuracy": 0.20625,
             "count": 320
           },
           "SS": {
-            "accuracy": 0.625,
+            "accuracy": 0.5357142857142857,
             "count": 56
           },
           "UC": {
-            "accuracy": 0.3534971644612476,
+            "accuracy": 0.3572778827977316,
             "count": 529
           },
           "US": {
-            "accuracy": 0.5,
+            "accuracy": 0.6458333333333334,
             "count": 48
           }
         }
@@ -982,7 +982,7 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.4,
+            "accuracy": 0.3933333333333333,
             "count": 300
           },
           "SC": {
@@ -990,7 +990,7 @@
             "count": 100
           },
           "UC": {
-            "accuracy": 0.22279792746113988,
+            "accuracy": 0.23316062176165803,
             "count": 193
           },
           "US": {
@@ -1004,41 +1004,41 @@
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.53,
+            "accuracy": 0.52,
             "count": 200
           },
           "SC": {
-            "accuracy": 0.04,
+            "accuracy": 0.05,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.19921875,
+            "accuracy": 0.21484375,
             "count": 256
           },
           "US": {
-            "accuracy": 0.4513888888888889,
+            "accuracy": 0.4097222222222222,
             "count": 144
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.01,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.57,
+            "accuracy": 0.53,
             "count": 100
           },
           "SC": {
-            "accuracy": 0.09,
+            "accuracy": 0.06,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.30392156862745096,
+            "accuracy": 0.2777777777777778,
             "count": 306
           },
           "US": {
-            "accuracy": 0.4948453608247423,
+            "accuracy": 0.5154639175257731,
             "count": 194
           }
         }
@@ -1048,15 +1048,15 @@
         "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.09,
+            "accuracy": 0.18,
             "count": 100
           },
           "UC": {
-            "accuracy": 0.31693989071038253,
+            "accuracy": 0.33879781420765026,
             "count": 366
           },
           "US": {
-            "accuracy": 0.7521367521367521,
+            "accuracy": 0.8632478632478633,
             "count": 234
           }
         }
@@ -1066,11 +1066,11 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.3277870216306156,
+            "accuracy": 0.33111480865224624,
             "count": 601
           },
           "ME": {
-            "accuracy": 0.9090909090909091,
+            "accuracy": 0.9696969696969697,
             "count": 99
           }
         }
@@ -1080,19 +1080,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.5483870967741935,
+            "accuracy": 0.5376344086021505,
             "count": 279
           },
           "MB": {
-            "accuracy": 0.05517241379310345,
+            "accuracy": 0.020689655172413793,
             "count": 145
           },
           "ME": {
-            "accuracy": 0.6666666666666666,
+            "accuracy": 0.9166666666666666,
             "count": 24
           },
           "UB": {
-            "accuracy": 0.2222222222222222,
+            "accuracy": 0.21428571428571427,
             "count": 252
           }
         }
@@ -1102,23 +1102,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.7183098591549296,
+            "accuracy": 0.7136150234741784,
             "count": 213
           },
           "MB": {
-            "accuracy": 0.017699115044247787,
+            "accuracy": 0.02654867256637168,
             "count": 113
           },
           "ME": {
-            "accuracy": 0.9058823529411765,
+            "accuracy": 0.9411764705882353,
             "count": 85
           },
           "UB": {
-            "accuracy": 0.23204419889502761,
+            "accuracy": 0.2265193370165746,
             "count": 181
           },
           "UD": {
-            "accuracy": 0.2037037037037037,
+            "accuracy": 0.10185185185185185,
             "count": 108
           }
         }
@@ -1128,23 +1128,23 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.7932960893854749,
+            "accuracy": 0.8156424581005587,
             "count": 179
           },
           "MB": {
-            "accuracy": 0.02912621359223301,
+            "accuracy": 0.0,
             "count": 103
           },
           "ME": {
-            "accuracy": 0.9285714285714286,
+            "accuracy": 0.9821428571428571,
             "count": 56
           },
           "UB": {
-            "accuracy": 0.33557046979865773,
+            "accuracy": 0.26174496644295303,
             "count": 149
           },
           "UD": {
-            "accuracy": 0.2112676056338028,
+            "accuracy": 0.07981220657276995,
             "count": 213
           }
         }
@@ -1154,25 +1154,25 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.725,
+            "accuracy": 0.755,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.1,
+            "accuracy": 0.01,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.42,
+            "accuracy": 0.38,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.31,
+            "accuracy": 0.09666666666666666,
             "count": 300
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.03,
+        "full_accuracy": 0.01,
         "n_examples": 100,
         "per_subtask": {
           "MD": {
@@ -1180,15 +1180,15 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.03,
+            "accuracy": 0.02,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.71,
+            "accuracy": 0.53,
             "count": 100
           },
           "UD": {
-            "accuracy": 0.285,
+            "accuracy": 0.11,
             "count": 400
           }
         }
@@ -1198,23 +1198,23 @@
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.53,
+            "accuracy": 0.5216666666666666,
             "count": 600
           },
           "MB": {
-            "accuracy": 0.03745318352059925,
+            "accuracy": 0.0149812734082397,
             "count": 267
           },
           "ME": {
-            "accuracy": 0.7358490566037735,
+            "accuracy": 0.9056603773584906,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.23917995444191345,
+            "accuracy": 0.2460136674259681,
             "count": 439
           },
           "UD": {
-            "accuracy": 0.17073170731707318,
+            "accuracy": 0.07317073170731707,
             "count": 41
           }
         }
@@ -1224,19 +1224,19 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.51,
+            "accuracy": 0.5266666666666666,
             "count": 300
           },
           "MB": {
-            "accuracy": 0.13,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.18274111675126903,
+            "accuracy": 0.2233502538071066,
             "count": 197
           },
           "UD": {
-            "accuracy": 0.39805825242718446,
+            "accuracy": 0.10679611650485436,
             "count": 103
           }
         }
@@ -1246,25 +1246,25 @@
         "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.665,
+            "accuracy": 0.655,
             "count": 200
           },
           "MB": {
-            "accuracy": 0.1,
+            "accuracy": 0.03,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.21862348178137653,
+            "accuracy": 0.24696356275303644,
             "count": 247
           },
           "UD": {
-            "accuracy": 0.38562091503267976,
+            "accuracy": 0.11764705882352941,
             "count": 153
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.02,
+        "full_accuracy": 0.0,
         "n_examples": 100,
         "per_subtask": {
           "MD": {
@@ -1272,26 +1272,26 @@
             "count": 100
           },
           "MB": {
-            "accuracy": 0.07,
+            "accuracy": 0.01,
             "count": 100
           },
           "UB": {
-            "accuracy": 0.2953020134228188,
+            "accuracy": 0.2651006711409396,
             "count": 298
           },
           "UD": {
-            "accuracy": 0.33663366336633666,
+            "accuracy": 0.16336633663366337,
             "count": 202
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.008333333333333333,
+      "overall_accuracy": 0.007083333333333333,
       "total_examples": 2400,
       "n_splits": 22
     }
   },
-  "sorl_overall_accuracy": 0.008333333333333333,
-  "sft_overall_accuracy": 0.004166666666666667
+  "sorl_overall_accuracy": 0.007083333333333333,
+  "sft_overall_accuracy": 0.007083333333333333
 }

add_sub_sorl_v1_abs10_10K/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5539570ac41c2434a03e526e01a1a87e8aa66fcfadd6151a66949f140ded418d
+oid sha256:324ce6c56d7206c9ce03c2a04c06b931963e334e14fddc37c56898ab88915205
 size 650303660

add_sub_sorl_v1_abs10_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,
@@ -69,16 +69,16 @@
   "no_wandb": false,
   "n_params": 162499262,
   "run_name": "add_sub_sorl_v1_abs10_10K",
-  "git_commit": "78d46f8665a87f4b44bd5894bd34f393f2dea51f",
-  "timestamp": "2026-04-12T08:59:09.036996+00:00",
+  "git_commit": "8d5ee5420119746ef4e2c87570eb250c9718f643",
+  "timestamp": "2026-04-12T22:18:34.743639+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": "f2ajfc34",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/f2ajfc34",
-  "final_accuracy": 0.008333333333333333,
-  "sft_accuracy": 0.004166666666666667,
+  "wandb_run_id": "hd7zspyv",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/hd7zspyv",
+  "final_accuracy": 0.007083333333333333,
+  "sft_accuracy": 0.007083333333333333,
   "eval_method": "ArithmeticEvaluator"
 }