fix f1_score calculation bug

external_models_results.json

@@ -14,12 +14,12 @@
             "assin2_sts": 0.7053302344881672,
             "assin2_rte": 0.9121728362223306,
             "faquad_nli": 0.7575848453041435,
-            "hatebr_offensive": 0.5025338637870607,
-            "portuguese_hate_speech": 0.4650217578860529,
-            "tweetsentbr": 0.533977453070735
+            "hatebr_offensive": 0.753800795680591,
+            "portuguese_hate_speech": 0.6975326368290793,
+            "tweetsentbr": 0.7119699374276466
         },
-        "result_metrics_average": 0.6428099652929031,
-        "result_metrics_npm": 0.43960062672137007
+        "result_metrics_average": 0.7163399980921773,
+        "result_metrics_npm": 0.5744541501392351
     },
     {
         "model": "sabia-2-medium",
@@ -82,10 +82,10 @@
             "faquad_nli": 0.6340996599941455,
             "hatebr_offensive": 0.8023698759439051,
             "portuguese_hate_speech": 0.7342166269560177,
-            "tweetsentbr": 0.5477486799750156
+            "tweetsentbr": 0.7303315733000207
         },
-        "result_metrics_average": 0.7212271446046878,
-        "result_metrics_npm": 0.5735261536314672
+        "result_metrics_average": 0.7415141327519107,
+        "result_metrics_npm": 0.6037151240886439
     },
     {
         "model": "gemini-1.0-pro",
@@ -122,14 +122,14 @@
             "bluex": 0.7719054242002782,
             "oab_exams": 0.6888382687927107,
             "assin2_sts": 0.8159702278408203,
-            "assin2_rte": 0.9328989988467518,
-            "faquad_nli": 0.7290756302521009,
-            "hatebr_offensive": 0.8697698647467024,
-            "portuguese_hate_speech": 0.7539414414414414,
-            "tweetsentbr": 0.772785080895884
+            "assin2_rte": 0.4651063829787234,
+            "faquad_nli": 0.4114285714285714,
+            "hatebr_offensive": 0.4422336328626444,
+            "portuguese_hate_speech": 0.41441441441441434,
+            "tweetsentbr": 0.7725066133902373
         },
-        "result_metrics_average": 0.7984588504001905,
-        "result_metrics_npm": 0.6908188311933006
+        "result_metrics_average": 0.6259275836103805,
+        "result_metrics_npm": 0.3545226679161642
     },
     {
         "model": "deepseek-v2-chat",
@@ -256,14 +256,14 @@
             "bluex": 0.8011126564673157,
             "oab_exams": 0.7640091116173121,
             "assin2_sts": 0.7888441732870783,
-            "assin2_rte": 0.6317630318610981,
+            "assin2_rte": 0.9476445477916471,
             "faquad_nli": 0.825063276593557,
             "hatebr_offensive": 0.9073940659389119,
             "portuguese_hate_speech": 0.7191480935512969,
             "tweetsentbr": 0.7821434639106575
         },
-        "result_metrics_average": 0.7857580188820238,
-        "result_metrics_npm": 0.6584973442501938
+        "result_metrics_average": 0.8208559650965292,
+        "result_metrics_npm": 0.7286932366792048
     },
     {
         "model": "sabia-3",
@@ -280,11 +280,11 @@
             "assin2_sts": 0.8253863689009022,
             "assin2_rte": 0.9477034821619312,
             "faquad_nli": 0.8243848812618203,
-            "hatebr_offensive": 0.5519158516393349,
-            "portuguese_hate_speech": 0.48273809523809524,
-            "tweetsentbr": 0.5632959814986498
+            "hatebr_offensive": 0.8278737774590023,
+            "portuguese_hate_speech": 0.7241071428571428,
+            "tweetsentbr": 0.7510613086648664
         },
-        "result_metrics_average": 0.744836336226786,
-        "result_metrics_npm": 0.5802643096708316
+        "result_metrics_average": 0.8231799251828895,
+        "result_metrics_npm": 0.7241097388486535
     }
 ]