Upload add_sub_sorl_v1_abs30_K1_10K/metrics.json with huggingface_hub

add_sub_sorl_v1_abs30_K1_10K/metrics.json

@@ -550,502 +550,567 @@
       "K": null,
       "mode": "sft",
       "n_digits": 6,
-      "n_per_split": 100
+      "n_per_split": 50
     },
     "splits": {
       "add_S0": {
-        "full_accuracy": 0.97,
-        "n_examples": 100,
+        "full_accuracy": 0.94,
+        "digit_accuracy": 0.9914285714285714,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
-            "accuracy": 0.9950413223140496,
-            "count": 605
+            "accuracy": 0.9898305084745763,
+            "count": 295
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 95
+            "count": 55
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 204
+            "count": 126
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 169
+            "count": 79
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 31
+            "count": 21
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 296
+            "count": 124
           }
         }
       },
       "add_S2": {
         "full_accuracy": 0.96,
-        "n_examples": 100,
+        "digit_accuracy": 0.9942857142857143,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 163
+            "count": 75
           },
           "SC": {
-            "accuracy": 0.9769230769230769,
-            "count": 130
+            "accuracy": 0.967741935483871,
+            "count": 62
           },
           "SS": {
-            "accuracy": 0.9770114942528736,
-            "count": 87
+            "accuracy": 1.0,
+            "count": 39
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 203
+            "count": 111
           },
           "US": {
             "accuracy": 1.0,
-            "count": 117
+            "count": 63
           }
         }
       },
       "add_S3": {
-        "full_accuracy": 0.64,
-        "n_examples": 100,
+        "full_accuracy": 0.68,
+        "digit_accuracy": 0.9542857142857143,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 121
+            "count": 60
           },
           "SC": {
-            "accuracy": 0.9834710743801653,
-            "count": 121
+            "accuracy": 0.9824561403508771,
+            "count": 57
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 49
+            "count": 19
           },
           "UC": {
-            "accuracy": 0.8172043010752689,
-            "count": 186
+            "accuracy": 0.8557692307692307,
+            "count": 104
           },
           "US": {
             "accuracy": 1.0,
-            "count": 223
+            "count": 110
           }
         }
       },
       "add_S4": {
-        "full_accuracy": 0.59,
-        "n_examples": 100,
+        "full_accuracy": 0.66,
+        "digit_accuracy": 0.94,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 104
+            "count": 48
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 106
+            "count": 52
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 23
+            "count": 7
           },
           "UC": {
-            "accuracy": 0.8125,
-            "count": 160
+            "accuracy": 0.898876404494382,
+            "count": 89
           },
           "US": {
-            "accuracy": 0.9153094462540716,
-            "count": 307
+            "accuracy": 0.922077922077922,
+            "count": 154
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.62,
-        "n_examples": 100,
+        "full_accuracy": 0.68,
+        "digit_accuracy": 0.8714285714285714,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.71,
-            "count": 100
+            "accuracy": 0.74,
+            "count": 50
           },
           "US": {
-            "accuracy": 0.86,
-            "count": 400
+            "accuracy": 0.84,
+            "count": 200
           }
         }
       },
       "add_S6": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "US": {
             "accuracy": 1.0,
-            "count": 500
+            "count": 250
           }
         }
       },
       "add_random": {
         "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 447
+            "count": 431
           },
           "SC": {
-            "accuracy": 0.996875,
-            "count": 320
+            "accuracy": 1.0,
+            "count": 316
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 39
           },
           "UC": {
-            "accuracy": 0.998109640831758,
-            "count": 529
+            "accuracy": 0.9964285714285714,
+            "count": 560
           },
           "US": {
             "accuracy": 1.0,
-            "count": 48
+            "count": 54
           }
         }
       },
-      "add_C3": {
-        "full_accuracy": 0.84,
-        "n_examples": 100,
+      "add_C1": {
+        "full_accuracy": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.996,
+            "count": 250
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 50
+          }
+        }
+      },
+      "add_C2": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 300
+            "count": 200
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.917098445595855,
-            "count": 193
+            "accuracy": 1.0,
+            "count": 83
           },
           "US": {
             "accuracy": 1.0,
-            "count": 107
+            "count": 17
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.8,
+        "digit_accuracy": 0.9714285714285714,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9933333333333333,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.91,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 50
           }
         }
       },
       "add_C4": {
-        "full_accuracy": 0.78,
-        "n_examples": 100,
+        "full_accuracy": 0.86,
+        "digit_accuracy": 0.9771428571428571,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.92578125,
-            "count": 256
+            "accuracy": 0.9545454545454546,
+            "count": 132
           },
           "US": {
-            "accuracy": 0.9583333333333334,
-            "count": 144
+            "accuracy": 0.9705882352941176,
+            "count": 68
           }
         }
       },
       "add_C5": {
-        "full_accuracy": 0.72,
-        "n_examples": 100,
+        "full_accuracy": 0.76,
+        "digit_accuracy": 0.9571428571428572,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.9313725490196079,
-            "count": 306
+            "accuracy": 0.9383561643835616,
+            "count": 146
           },
           "US": {
-            "accuracy": 0.9536082474226805,
-            "count": 194
+            "accuracy": 0.9423076923076923,
+            "count": 104
           }
         }
       },
       "add_C6": {
-        "full_accuracy": 0.82,
-        "n_examples": 100,
+        "full_accuracy": 0.92,
+        "digit_accuracy": 0.9885714285714285,
+        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.953551912568306,
-            "count": 366
+            "accuracy": 0.9788359788359788,
+            "count": 189
           },
           "US": {
-            "accuracy": 0.9957264957264957,
-            "count": 234
+            "accuracy": 1.0,
+            "count": 111
           }
         }
       },
       "sub_M0": {
-        "full_accuracy": 0.96,
-        "n_examples": 100,
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
-            "accuracy": 0.9933444259567388,
-            "count": 601
+            "accuracy": 1.0,
+            "count": 303
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 99
+            "count": 47
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 279
+            "count": 141
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 145
+            "count": 72
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 24
+            "count": 18
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 252
+            "count": 119
           }
         }
       },
       "sub_M2": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 213
+            "count": 112
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 113
+            "count": 53
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 85
+            "count": 47
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 181
+            "count": 85
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 108
+            "count": 53
           }
         }
       },
       "sub_M3": {
-        "full_accuracy": 0.35,
-        "n_examples": 100,
+        "full_accuracy": 0.32,
+        "digit_accuracy": 0.9028571428571428,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 179
+            "count": 97
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 103
+            "count": 51
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 27
           },
           "UB": {
-            "accuracy": 0.5637583892617449,
-            "count": 149
+            "accuracy": 0.5405405405405406,
+            "count": 74
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 213
+            "count": 101
           }
         }
       },
       "sub_M4": {
-        "full_accuracy": 0.03,
-        "n_examples": 100,
+        "full_accuracy": 0.0,
+        "digit_accuracy": 0.8028571428571428,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.49,
-            "count": 100
+            "accuracy": 0.6,
+            "count": 50
           },
           "UD": {
-            "accuracy": 0.6766666666666666,
-            "count": 300
+            "accuracy": 0.6733333333333333,
+            "count": 150
           }
         }
       },
       "sub_M5": {
         "full_accuracy": 0.0,
-        "n_examples": 100,
+        "digit_accuracy": 0.6428571428571429,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.38,
-            "count": 100
+            "accuracy": 0.46,
+            "count": 50
           },
           "UD": {
-            "accuracy": 0.4975,
-            "count": 400
+            "accuracy": 0.51,
+            "count": 200
           }
         }
       },
       "sub_random": {
-        "full_accuracy": 0.995,
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 600
+            "count": 570
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 267
+            "count": 277
           },
           "ME": {
             "accuracy": 1.0,
             "count": 53
           },
           "UB": {
-            "accuracy": 0.9977220956719818,
-            "count": 439
+            "accuracy": 1.0,
+            "count": 471
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 41
+            "count": 29
           }
         }
       },
       "sub_B3": {
-        "full_accuracy": 0.84,
-        "n_examples": 100,
+        "full_accuracy": 0.9,
+        "digit_accuracy": 0.9857142857142858,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 300
+            "count": 150
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.9187817258883249,
-            "count": 197
+            "accuracy": 0.9504950495049505,
+            "count": 101
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 103
+            "count": 49
           }
         }
       },
       "sub_B4": {
-        "full_accuracy": 0.76,
-        "n_examples": 100,
+        "full_accuracy": 0.68,
+        "digit_accuracy": 0.9457142857142857,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.9311740890688259,
-            "count": 247
+            "accuracy": 0.9008264462809917,
+            "count": 121
           },
           "UD": {
-            "accuracy": 0.9084967320261438,
-            "count": 153
+            "accuracy": 0.9113924050632911,
+            "count": 79
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.59,
-        "n_examples": 100,
+        "full_accuracy": 0.74,
+        "digit_accuracy": 0.9514285714285714,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.8859060402684564,
-            "count": 298
+            "accuracy": 0.9407894736842105,
+            "count": 152
           },
           "UD": {
-            "accuracy": 0.8861386138613861,
-            "count": 202
+            "accuracy": 0.9183673469387755,
+            "count": 98
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.7683333333333333,
-      "total_examples": 2400,
-      "n_splits": 22
+      "overall_accuracy": 0.828,
+      "digit_accuracy": 0.9622857142857143,
+      "total_examples": 1500,
+      "n_splits": 24
     }
   },
   "sorl_eval": {
@@ -1054,416 +1119,477 @@
       "K": 1,
       "mode": "sorl",
       "n_digits": 6,
-      "n_per_split": 100
+      "n_per_split": 50
     },
     "splits": {
       "add_S0": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 605
+            "count": 295
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 95
+            "count": 55
           }
         }
       },
       "add_S1": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 204
+            "count": 126
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 169
+            "count": 79
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 31
+            "count": 21
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 296
+            "count": 124
           }
         }
       },
       "add_S2": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 163
+            "count": 75
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 130
+            "count": 62
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 87
+            "count": 39
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 203
+            "count": 111
           },
           "US": {
             "accuracy": 1.0,
-            "count": 117
+            "count": 63
           }
         }
       },
       "add_S3": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 121
+            "count": 60
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 121
+            "count": 57
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 49
+            "count": 19
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 186
+            "count": 104
           },
           "US": {
             "accuracy": 1.0,
-            "count": 223
+            "count": 110
           }
         }
       },
       "add_S4": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 104
+            "count": 48
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 106
+            "count": 52
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 23
+            "count": 7
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 160
+            "count": 89
           },
           "US": {
             "accuracy": 1.0,
-            "count": 307
+            "count": 154
           }
         }
       },
       "add_S5": {
-        "full_accuracy": 0.99,
-        "n_examples": 100,
+        "full_accuracy": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
-            "accuracy": 0.99,
-            "count": 100
+            "accuracy": 0.98,
+            "count": 50
           },
           "US": {
             "accuracy": 1.0,
-            "count": 400
+            "count": 200
           }
         }
       },
       "add_S6": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "US": {
             "accuracy": 1.0,
-            "count": 500
+            "count": 250
           }
         }
       },
       "add_random": {
         "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
         "n_examples": 200,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 447
+            "count": 431
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 320
+            "count": 316
           },
           "SS": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 39
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 529
+            "count": 560
           },
           "US": {
             "accuracy": 1.0,
-            "count": 48
+            "count": 54
+          }
+        }
+      },
+      "add_C1": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 250
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 50
+          }
+        }
+      },
+      "add_C2": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 83
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 17
           }
         }
       },
       "add_C3": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 300
+            "count": 150
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 193
+            "count": 100
           },
           "US": {
             "accuracy": 1.0,
-            "count": 107
+            "count": 50
           }
         }
       },
       "add_C4": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 256
+            "count": 132
           },
           "US": {
             "accuracy": 1.0,
-            "count": 144
+            "count": 68
           }
         }
       },
       "add_C5": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SA": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 306
+            "count": 146
           },
           "US": {
             "accuracy": 1.0,
-            "count": 194
+            "count": 104
           }
         }
       },
       "add_C6": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "SC": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UC": {
             "accuracy": 1.0,
-            "count": 366
+            "count": 189
           },
           "US": {
             "accuracy": 1.0,
-            "count": 234
+            "count": 111
           }
         }
       },
       "sub_M0": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 601
+            "count": 303
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 99
+            "count": 47
           }
         }
       },
       "sub_M1": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 279
+            "count": 141
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 145
+            "count": 72
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 24
+            "count": 18
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 252
+            "count": 119
           }
         }
       },
       "sub_M2": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 213
+            "count": 112
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 113
+            "count": 53
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 85
+            "count": 47
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 181
+            "count": 85
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 108
+            "count": 53
           }
         }
       },
       "sub_M3": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 179
+            "count": 97
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 103
+            "count": 51
           },
           "ME": {
             "accuracy": 1.0,
-            "count": 56
+            "count": 27
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 149
+            "count": 74
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 213
+            "count": 101
           }
         }
       },
       "sub_M4": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 300
+            "count": 150
           }
         }
       },
       "sub_M5": {
-        "full_accuracy": 0.14,
-        "n_examples": 100,
+        "full_accuracy": 0.06,
+        "digit_accuracy": 0.8657142857142858,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.14,
-            "count": 100
+            "accuracy": 0.06,
+            "count": 50
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 400
+            "count": 200
           }
         }
       },
       "sub_random": {
         "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
         "n_examples": 200,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 600
+            "count": 570
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 267
+            "count": 277
           },
           "ME": {
             "accuracy": 1.0,
@@ -1471,85 +1597,89 @@
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 439
+            "count": 471
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 41
+            "count": 29
           }
         }
       },
       "sub_B3": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 300
+            "count": 150
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 197
+            "count": 101
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 103
+            "count": 49
           }
         }
       },
       "sub_B4": {
         "full_accuracy": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 200
+            "count": 100
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 247
+            "count": 121
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 153
+            "count": 79
           }
         }
       },
       "sub_B5": {
-        "full_accuracy": 0.94,
-        "n_examples": 100,
+        "full_accuracy": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
-            "accuracy": 0.9798657718120806,
-            "count": 298
+            "accuracy": 0.993421052631579,
+            "count": 152
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 202
+            "count": 98
           }
         }
       }
     },
     "summary": {
-      "overall_accuracy": 0.96125,
-      "total_examples": 2400,
-      "n_splits": 22
+      "overall_accuracy": 0.9673333333333334,
+      "digit_accuracy": 0.9953333333333333,
+      "total_examples": 1500,
+      "n_splits": 24
     }
   },
   "sorl_overall_accuracy": 0.96125,