Upload add_sub_sorl_v1_abs30_100K/metrics.json with huggingface_hub

add_sub_sorl_v1_abs30_100K/metrics.json

@@ -5030,416 +5030,477 @@
       "K": null,
       "mode": "sft",
       "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": 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": 100
+            "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": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 100
+            "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,
@@ -5447,85 +5508,89 @@
           },
           "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": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 298
+            "count": 152
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 202
+            "count": 98
           }
         }
       }
     },
     "summary": {
       "overall_accuracy": 1.0,
-      "total_examples": 2400,
-      "n_splits": 22
+      "digit_accuracy": 1.0,
+      "total_examples": 1500,
+      "n_splits": 24
     }
   },
   "sorl_eval": {
@@ -5534,416 +5599,477 @@
       "K": 4,
       "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": 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": 100
+            "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": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 100
+            "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,
@@ -5951,85 +6077,89 @@
           },
           "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": 1.0,
-        "n_examples": 100,
+        "digit_accuracy": 1.0,
+        "n_examples": 50,
         "per_subtask": {
           "MD": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "MB": {
             "accuracy": 1.0,
-            "count": 100
+            "count": 50
           },
           "UB": {
             "accuracy": 1.0,
-            "count": 298
+            "count": 152
           },
           "UD": {
             "accuracy": 1.0,
-            "count": 202
+            "count": 98
           }
         }
       }
     },
     "summary": {
       "overall_accuracy": 1.0,
-      "total_examples": 2400,
-      "n_splits": 22
+      "digit_accuracy": 1.0,
+      "total_examples": 1500,
+      "n_splits": 24
     }
   },
   "sorl_overall_accuracy": 1.0,