Upload add_sub_baseline_10K_2L1H128d

add_sub_baseline_10K_2L1H128d/metrics.json

@@ -65,132 +65,132 @@
       3100
     ],
     "loss": [
-      11.632269859313965,
-      10.746880531311035,
-      10.444467544555664,
-      10.006251335144043,
-      9.801725387573242,
-      9.507078170776367,
-      9.177903175354004,
-      8.873117446899414,
-      8.627289772033691,
-      8.425443649291992,
-      8.176294326782227,
-      7.824686527252197,
-      7.672322750091553,
-      7.39719820022583,
-      7.087851524353027,
-      6.878063678741455,
-      6.657628059387207,
-      6.365849018096924,
-      6.228781223297119,
-      5.940927982330322,
-      5.686161994934082,
-      5.552325248718262,
-      5.286587715148926,
-      5.14711856842041,
-      4.970963001251221,
-      4.826735019683838,
-      4.654284954071045,
-      4.47059965133667,
-      4.298427581787109,
-      4.127253532409668,
-      4.005905628204346,
-      3.8805015087127686,
-      3.801790475845337,
-      3.6226108074188232,
-      3.5523719787597656,
-      3.5552258491516113,
-      3.372868061065674,
-      3.3437774181365967,
-      3.2328782081604004,
-      3.155165910720825,
-      3.211038112640381,
-      3.113257884979248,
-      2.977780818939209,
-      3.0747323036193848,
-      2.98803448677063,
-      2.9421255588531494,
-      2.9898440837860107,
-      2.9196650981903076,
-      2.8956546783447266,
-      2.9016549587249756,
-      2.8742430210113525,
-      2.872373342514038,
-      2.8761067390441895,
-      2.807882070541382,
-      2.8126957416534424,
-      2.813640594482422,
-      2.8270370960235596,
-      2.8576278686523438,
-      2.861673593521118,
-      2.809460401535034,
-      2.736629009246826,
-      2.832703113555908
+      11.642115592956543,
+      10.80324649810791,
+      10.34417724609375,
+      10.04995059967041,
+      9.737959861755371,
+      9.45477294921875,
+      9.285505294799805,
+      8.961015701293945,
+      8.579405784606934,
+      8.42731761932373,
+      8.175429344177246,
+      7.936665058135986,
+      7.630360126495361,
+      7.336586952209473,
+      7.065025329589844,
+      6.900625705718994,
+      6.648900508880615,
+      6.47622013092041,
+      6.186995029449463,
+      5.953526020050049,
+      5.7461323738098145,
+      5.58115291595459,
+      5.248443603515625,
+      5.121757507324219,
+      4.968733787536621,
+      4.763223171234131,
+      4.608892917633057,
+      4.488630771636963,
+      4.2728352546691895,
+      4.094915866851807,
+      4.042762756347656,
+      3.8673317432403564,
+      3.705319881439209,
+      3.728566884994507,
+      3.5413622856140137,
+      3.5115506649017334,
+      3.3972582817077637,
+      3.352463483810425,
+      3.3156659603118896,
+      3.187150478363037,
+      3.145035982131958,
+      3.1074206829071045,
+      3.0765254497528076,
+      3.0211308002471924,
+      2.9677698612213135,
+      2.983156204223633,
+      2.895720958709717,
+      2.862177610397339,
+      2.9088594913482666,
+      2.9126336574554443,
+      2.8361401557922363,
+      2.851780414581299,
+      2.870859384536743,
+      2.7742607593536377,
+      2.7949109077453613,
+      2.836515426635742,
+      2.8090367317199707,
+      2.846834421157837,
+      2.775794506072998,
+      2.7425057888031006,
+      2.8071517944335938,
+      2.822736978530884
     ],
     "base_loss": [
-      11.632269859313965,
-      10.746880531311035,
-      10.444467544555664,
-      10.006251335144043,
-      9.801725387573242,
-      9.507078170776367,
-      9.177903175354004,
-      8.873117446899414,
-      8.627289772033691,
-      8.425443649291992,
-      8.176294326782227,
-      7.824686527252197,
-      7.672322750091553,
-      7.39719820022583,
-      7.087851524353027,
-      6.878063678741455,
-      6.657628059387207,
-      6.365849018096924,
-      6.228781223297119,
-      5.940927982330322,
-      5.686161994934082,
-      5.552325248718262,
-      5.286587715148926,
-      5.14711856842041,
-      4.970963001251221,
-      4.826735019683838,
-      4.654284954071045,
-      4.47059965133667,
-      4.298427581787109,
-      4.127253532409668,
-      4.005905628204346,
-      3.8805015087127686,
-      3.801790475845337,
-      3.6226108074188232,
-      3.5523719787597656,
-      3.5552258491516113,
-      3.372868061065674,
-      3.3437774181365967,
-      3.2328782081604004,
-      3.155165910720825,
-      3.211038112640381,
-      3.113257884979248,
-      2.977780818939209,
-      3.0747323036193848,
-      2.98803448677063,
-      2.9421255588531494,
-      2.9898440837860107,
-      2.9196650981903076,
-      2.8956546783447266,
-      2.9016549587249756,
-      2.8742430210113525,
-      2.872373342514038,
-      2.8761067390441895,
-      2.807882070541382,
-      2.8126957416534424,
-      2.813640594482422,
-      2.8270370960235596,
-      2.8576278686523438,
-      2.861673593521118,
-      2.809460401535034,
-      2.736629009246826,
-      2.832703113555908
+      11.642115592956543,
+      10.80324649810791,
+      10.34417724609375,
+      10.04995059967041,
+      9.737959861755371,
+      9.45477294921875,
+      9.285505294799805,
+      8.961015701293945,
+      8.579405784606934,
+      8.42731761932373,
+      8.175429344177246,
+      7.936665058135986,
+      7.630360126495361,
+      7.336586952209473,
+      7.065025329589844,
+      6.900625705718994,
+      6.648900508880615,
+      6.47622013092041,
+      6.186995029449463,
+      5.953526020050049,
+      5.7461323738098145,
+      5.58115291595459,
+      5.248443603515625,
+      5.121757507324219,
+      4.968733787536621,
+      4.763223171234131,
+      4.608892917633057,
+      4.488630771636963,
+      4.2728352546691895,
+      4.094915866851807,
+      4.042762756347656,
+      3.8673317432403564,
+      3.705319881439209,
+      3.728566884994507,
+      3.5413622856140137,
+      3.5115506649017334,
+      3.3972582817077637,
+      3.352463483810425,
+      3.3156659603118896,
+      3.187150478363037,
+      3.145035982131958,
+      3.1074206829071045,
+      3.0765254497528076,
+      3.0211308002471924,
+      2.9677698612213135,
+      2.983156204223633,
+      2.895720958709717,
+      2.862177610397339,
+      2.9088594913482666,
+      2.9126336574554443,
+      2.8361401557922363,
+      2.851780414581299,
+      2.870859384536743,
+      2.7742607593536377,
+      2.7949109077453613,
+      2.836515426635742,
+      2.8090367317199707,
+      2.846834421157837,
+      2.775794506072998,
+      2.7425057888031006,
+      2.8071517944335938,
+      2.822736978530884
     ],
     "lr": [
       9.800000000000001e-06,
@@ -330,167 +330,167 @@
       "K": null,
       "mode": "sft",
       "n_digits": 6,
-      "n_per_split": 50
+      "n_per_split": 100
     },
     "splits": {
       "add_S0": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.14857142857142858,
-        "n_examples": 50,
+        "digit_accuracy": 0.15428571428571428,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.17627118644067796,
-            "count": 295
+            "accuracy": 0.17851239669421487,
+            "count": 605
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 55
+            "count": 95
           }
         }
       },
       "add_S1": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.1657142857142857,
-        "n_examples": 50,
+        "digit_accuracy": 0.12857142857142856,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.30158730158730157,
-            "count": 126
+            "accuracy": 0.24509803921568626,
+            "count": 204
           },
           "SC": {
-            "accuracy": 0.25316455696202533,
-            "count": 79
+            "accuracy": 0.22485207100591717,
+            "count": 169
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 21
+            "count": 31
           },
           "UC": {
-            "accuracy": 0.0,
-            "count": 124
+            "accuracy": 0.006756756756756757,
+            "count": 296
           }
         }
       },
       "add_S2": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.28285714285714286,
-        "n_examples": 50,
+        "digit_accuracy": 0.26142857142857145,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.24,
-            "count": 75
+            "accuracy": 0.2331288343558282,
+            "count": 163
           },
           "SC": {
-            "accuracy": 0.2903225806451613,
-            "count": 62
+            "accuracy": 0.2,
+            "count": 130
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 39
+            "count": 87
           },
           "UC": {
-            "accuracy": 0.0,
-            "count": 111
+            "accuracy": 0.009852216748768473,
+            "count": 203
           },
           "US": {
             "accuracy": 1.0,
-            "count": 63
+            "count": 117
           }
         }
       },
       "add_S3": {
         "full_accuracy": 0.0,
         "digit_accuracy": 0.3942857142857143,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.26666666666666666,
-            "count": 60
+            "accuracy": 0.2727272727272727,
+            "count": 121
           },
           "SC": {
-            "accuracy": 0.21052631578947367,
-            "count": 57
+            "accuracy": 0.1652892561983471,
+            "count": 121
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 19
+            "count": 49
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 104
+            "count": 186
           },
           "US": {
             "accuracy": 1.0,
-            "count": 110
+            "count": 223
           }
         }
       },
       "add_S4": {
         "full_accuracy": 0.0,
         "digit_accuracy": 0.5085714285714286,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.20833333333333334,
-            "count": 48
+            "accuracy": 0.2403846153846154,
+            "count": 104
           },
           "SC": {
-            "accuracy": 0.2692307692307692,
-            "count": 52
+            "accuracy": 0.22641509433962265,
+            "count": 106
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 7
+            "count": 23
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 89
+            "count": 160
           },
           "US": {
             "accuracy": 1.0,
-            "count": 154
+            "count": 307
           }
         }
       },
       "add_S5": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.7,
-        "n_examples": 50,
+        "digit_accuracy": 0.6828571428571428,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.54,
-            "count": 50
+            "accuracy": 0.5,
+            "count": 100
           },
           "SC": {
-            "accuracy": 0.36,
-            "count": 50
+            "accuracy": 0.28,
+            "count": 100
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "US": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 400
           }
         }
       },
       "add_S6": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.7628571428571429,
-        "n_examples": 50,
+        "digit_accuracy": 0.7442857142857143,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.34,
-            "count": 50
+            "accuracy": 0.21,
+            "count": 100
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "US": {
             "accuracy": 1.0,
-            "count": 250
+            "count": 500
           }
         }
       },
@@ -500,292 +500,292 @@
         "n_examples": 200,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.23665893271461716,
-            "count": 431
+            "accuracy": 0.2371364653243848,
+            "count": 447
           },
           "SC": {
-            "accuracy": 0.1962025316455696,
-            "count": 316
+            "accuracy": 0.2125,
+            "count": 320
           },
           "SS": {
             "accuracy": 0.0,
-            "count": 39
+            "count": 56
           },
           "UC": {
-            "accuracy": 0.007142857142857143,
-            "count": 560
+            "accuracy": 0.0,
+            "count": 529
           },
           "US": {
             "accuracy": 1.0,
-            "count": 54
+            "count": 48
           }
         }
       },
       "add_C1": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.22857142857142856,
-        "n_examples": 50,
+        "digit_accuracy": 0.22428571428571428,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.272,
-            "count": 250
+            "accuracy": 0.262,
+            "count": 500
           },
           "SC": {
-            "accuracy": 0.24,
-            "count": 50
+            "accuracy": 0.26,
+            "count": 100
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           }
         }
       },
       "add_C2": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.26571428571428574,
-        "n_examples": 50,
+        "digit_accuracy": 0.26142857142857145,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.32,
-            "count": 200
+            "accuracy": 0.3025,
+            "count": 400
           },
           "SC": {
-            "accuracy": 0.22,
-            "count": 50
+            "accuracy": 0.18,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.012048192771084338,
-            "count": 83
+            "accuracy": 0.0,
+            "count": 156
           },
           "US": {
             "accuracy": 1.0,
-            "count": 17
+            "count": 44
           }
         }
       },
       "add_C3": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.34285714285714286,
-        "n_examples": 50,
+        "digit_accuracy": 0.32571428571428573,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.37333333333333335,
-            "count": 150
+            "accuracy": 0.3333333333333333,
+            "count": 300
           },
           "SC": {
-            "accuracy": 0.28,
-            "count": 50
+            "accuracy": 0.26,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.0,
-            "count": 100
+            "accuracy": 0.005025125628140704,
+            "count": 199
           },
           "US": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 101
           }
         }
       },
       "add_C4": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.3457142857142857,
-        "n_examples": 50,
+        "digit_accuracy": 0.3485714285714286,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.44,
-            "count": 100
+            "accuracy": 0.42,
+            "count": 200
           },
           "SC": {
-            "accuracy": 0.18,
-            "count": 50
+            "accuracy": 0.24,
+            "count": 100
           },
           "UC": {
             "accuracy": 0.0,
-            "count": 132
+            "count": 264
           },
           "US": {
             "accuracy": 1.0,
-            "count": 68
+            "count": 136
           }
         }
       },
       "add_C5": {
         "full_accuracy": 0.0,
         "digit_accuracy": 0.38857142857142857,
-        "n_examples": 50,
+        "n_examples": 100,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.48,
-            "count": 50
+            "accuracy": 0.59,
+            "count": 100
           },
           "SC": {
-            "accuracy": 0.16,
-            "count": 50
+            "accuracy": 0.22,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.0,
-            "count": 146
+            "accuracy": 0.0032258064516129032,
+            "count": 310
           },
           "US": {
             "accuracy": 1.0,
-            "count": 104
+            "count": 190
           }
         }
       },
       "add_C6": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.3514285714285714,
-        "n_examples": 50,
+        "digit_accuracy": 0.3585714285714286,
+        "n_examples": 100,
         "per_subtask": {
           "SC": {
-            "accuracy": 0.22,
-            "count": 50
+            "accuracy": 0.2,
+            "count": 100
           },
           "UC": {
-            "accuracy": 0.005291005291005291,
-            "count": 189
+            "accuracy": 0.002702702702702703,
+            "count": 370
           },
           "US": {
             "accuracy": 1.0,
-            "count": 111
+            "count": 230
           }
         }
       },
       "sub_M0": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.3057142857142857,
-        "n_examples": 50,
+        "digit_accuracy": 0.29285714285714287,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.19801980198019803,
-            "count": 303
+            "accuracy": 0.1951219512195122,
+            "count": 615
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 47
+            "count": 85
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.26285714285714284,
-        "n_examples": 50,
+        "digit_accuracy": 0.22428571428571428,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.3900709219858156,
-            "count": 141
+            "accuracy": 0.3698630136986301,
+            "count": 292
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 72
+            "count": 144
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 18
+            "count": 25
           },
           "UB": {
-            "accuracy": 0.15966386554621848,
-            "count": 119
+            "accuracy": 0.100418410041841,
+            "count": 239
           }
         }
       },
       "sub_M2": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.38285714285714284,
-        "n_examples": 50,
+        "digit_accuracy": 0.35428571428571426,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6428571428571429,
-            "count": 112
+            "accuracy": 0.6208530805687204,
+            "count": 211
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 53
+            "count": 115
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 47
+            "count": 85
           },
           "UB": {
-            "accuracy": 0.17647058823529413,
-            "count": 85
+            "accuracy": 0.17679558011049723,
+            "count": 181
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 53
+            "count": 108
           }
         }
       },
       "sub_M3": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.28,
-        "n_examples": 50,
+        "digit_accuracy": 0.3,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.6494845360824743,
-            "count": 97
+            "accuracy": 0.7597765363128491,
+            "count": 179
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 51
+            "count": 103
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 27
+            "count": 56
           },
           "UB": {
-            "accuracy": 0.10810810810810811,
-            "count": 74
+            "accuracy": 0.12080536912751678,
+            "count": 149
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 101
+            "count": 213
           }
         }
       },
       "sub_M4": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.21142857142857144,
-        "n_examples": 50,
+        "digit_accuracy": 0.18571428571428572,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 0.5,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.48,
-            "count": 50
+            "accuracy": 0.3,
+            "count": 100
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 150
+            "count": 300
           }
         }
       },
       "sub_M5": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.18285714285714286,
-        "n_examples": 50,
+        "digit_accuracy": 0.18714285714285714,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.28,
-            "count": 50
+            "accuracy": 0.31,
+            "count": 100
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 200
+            "count": 400
           }
         }
       },
@@ -795,101 +795,101 @@
         "n_examples": 200,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.37719298245614036,
-            "count": 570
+            "accuracy": 0.3616666666666667,
+            "count": 600
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 277
+            "count": 267
           },
           "ME": {
             "accuracy": 1.0,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.11677282377919321,
-            "count": 471
+            "accuracy": 0.12072892938496584,
+            "count": 439
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 29
+            "count": 41
           }
         }
       },
       "sub_B3": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.18857142857142858,
-        "n_examples": 50,
+        "digit_accuracy": 0.19285714285714287,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 0.3333333333333333,
-            "count": 150
+            "count": 300
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.15841584158415842,
-            "count": 101
+            "accuracy": 0.17766497461928935,
+            "count": 197
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 49
+            "count": 103
           }
         }
       },
       "sub_B4": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.17714285714285713,
-        "n_examples": 50,
+        "digit_accuracy": 0.19428571428571428,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 0.5,
-            "count": 100
+            "count": 200
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.09917355371900827,
-            "count": 121
+            "accuracy": 0.145748987854251,
+            "count": 247
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 79
+            "count": 153
           }
         }
       },
       "sub_B5": {
         "full_accuracy": 0.0,
-        "digit_accuracy": 0.18,
-        "n_examples": 50,
+        "digit_accuracy": 0.19,
+        "n_examples": 100,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 50
+            "count": 100
           },
           "MB": {
             "accuracy": 0.0,
-            "count": 50
+            "count": 100
           },
           "UB": {
-            "accuracy": 0.08552631578947369,
-            "count": 152
+            "accuracy": 0.11073825503355705,
+            "count": 298
           },
           "UD": {
             "accuracy": 0.0,
-            "count": 98
+            "count": 202
           }
         }
       }
     },
     "summary": {
       "overall_accuracy": 0.0,
-      "digit_accuracy": 0.2867619047619048,
-      "total_examples": 1500,
+      "digit_accuracy": 0.29516483516483516,
+      "total_examples": 2600,
       "n_splits": 24
     }
   }

add_sub_baseline_10K_2L1H128d/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7ed798fb07cbccdc55dcd6c1c4e872d058f4b5c532f1cd70f6a60ef685417772
+oid sha256:8a71d76648b9a847738fd64980f8c9b4abe52d0c748af9c58ccaa59d116373e7
 size 157692826

add_sub_baseline_10K_2L1H128d/train_config.json

@@ -69,15 +69,19 @@
   "no_wandb": false,
   "n_params": 39346560,
   "run_name": "add_sub_baseline_10K_2L1H128d",
-  "git_commit": "17e935f460a7f9595b705c1d614101a6b0e520f7",
-  "timestamp": "2026-04-14T06:22:14.025912+00:00",
+  "git_commit": "f835493c19eb98267697007042c9d440cad2afbb",
+  "timestamp": "2026-04-15T11:36:11.554711+00:00",
   "tokenizer": "Qwen/Qwen3-0.6B",
   "dataset_repo": "thoughtworks/arithmetic-sorl-data",
   "dataset_config": "add_sub_6digit",
+  "train_dataset": "fixed_train/train_10K_seed42.pt",
   "model_repo": "thoughtworks/arithmetic-sorl",
   "trainer_version": "sft",
-  "wandb_run_id": "lhp7f41c",
-  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/lhp7f41c",
+  "wandb_run_id": "zshwdb6w",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/zshwdb6w",
+  "eval_final_dataset": "eval_sets/eval_add_sub_6d_N100_seed42.json",
+  "eval_epoch_dataset": "eval_sets/eval_add_sub_6d_N25_seed42.json",
+  "eval_hf_repo": "thoughtworks/arithmetic-sorl-data",
   "final_accuracy": 0.0,
   "sft_accuracy": 0.0,
   "eval_method": "ArithmeticEvaluator"