[Run 20260606_135237] Log metrics

metrics/run_20260606_135237_metrics.json

@@ -0,0 +1,250 @@
+{
+  "run_id": "20260606_135237",
+  "hyperparameters": {
+    "model_architecture": "UBC-NLP/MARBERT",
+    "task": "sequence-classification",
+    "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
+    "learning_rate": 2e-05,
+    "batch_size": 16,
+    "num_epochs": 5,
+    "max_seq_length": 128,
+    "early_stopping_patience": 2,
+    "optimizer": "AdamW",
+    "seed": 42
+  },
+  "per_epoch_results": [
+    {
+      "epoch": 1,
+      "train_loss": 0.17704563700628081,
+      "val_loss": 0.08185144391769142,
+      "val_accuracy": 0.9764417177914111,
+      "val_f1_weighted": 0.9764376758355681,
+      "val_f1_macro": 0.9763831094316877,
+      "val_precision_macro": 0.976558193308877,
+      "val_recall_macro": 0.9762312443089591,
+      "hate_recall": 0.9716932578486875,
+      "hate_f1": 0.9752066115702479,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9787454639709694,
+          "recall": 0.9716932578486875,
+          "f1-score": 0.9752066115702479,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9743709226467847,
+          "recall": 0.9807692307692307,
+          "f1-score": 0.9775596072931276,
+          "support": 2132.0
+        },
+        "accuracy": 0.9764417177914111,
+        "macro avg": {
+          "precision": 0.976558193308877,
+          "recall": 0.9762312443089591,
+          "f1-score": 0.9763831094316877,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9764567468904389,
+          "recall": 0.9764417177914111,
+          "f1-score": 0.9764376758355681,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.08605310319347592,
+      "val_loss": 0.08948595821321029,
+      "val_accuracy": 0.9784049079754601,
+      "val_f1_weighted": 0.9783943419684894,
+      "val_f1_macro": 0.9783388704318936,
+      "val_precision_macro": 0.9788841169743632,
+      "val_recall_macro": 0.9779477298118323,
+      "hate_recall": 0.9680905815748841,
+      "hate_f1": 0.9771428571428571,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9863660199265862,
+          "recall": 0.9680905815748841,
+          "f1-score": 0.9771428571428571,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9714022140221402,
+          "recall": 0.9878048780487805,
+          "f1-score": 0.9795348837209302,
+          "support": 2132.0
+        },
+        "accuracy": 0.9784049079754601,
+        "macro avg": {
+          "precision": 0.9788841169743632,
+          "recall": 0.9779477298118323,
+          "f1-score": 0.9783388704318936,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9785371035613645,
+          "recall": 0.9784049079754601,
+          "f1-score": 0.9783943419684894,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.0567464292272778,
+      "val_loss": 0.09008082439386807,
+      "val_accuracy": 0.9820858895705521,
+      "val_f1_weighted": 0.982085266800533,
+      "val_f1_macro": 0.9820460322893185,
+      "val_precision_macro": 0.9820800840194114,
+      "val_recall_macro": 0.9820130520973447,
+      "hate_recall": 0.9804426145136387,
+      "hate_f1": 0.9812001030131342,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9819587628865979,
+          "recall": 0.9804426145136387,
+          "f1-score": 0.9812001030131342,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9822014051522249,
+          "recall": 0.9835834896810507,
+          "f1-score": 0.9828919615655027,
+          "support": 2132.0
+        },
+        "accuracy": 0.9820858895705521,
+        "macro avg": {
+          "precision": 0.9820800840194114,
+          "recall": 0.9820130520973447,
+          "f1-score": 0.9820460322893185,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9820857109382093,
+          "recall": 0.9820858895705521,
+          "f1-score": 0.982085266800533,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.03903503223161502,
+      "val_loss": 0.10641252933005578,
+      "val_accuracy": 0.978159509202454,
+      "val_f1_weighted": 0.9781645391606395,
+      "val_f1_macro": 0.9781232060259852,
+      "val_precision_macro": 0.9779342709417133,
+      "val_recall_macro": 0.9783747690994469,
+      "hate_recall": 0.9830159547092125,
+      "hate_f1": 0.9772320286518291,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9715157680569685,
+          "recall": 0.9830159547092125,
+          "f1-score": 0.9772320286518291,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.984352773826458,
+          "recall": 0.9737335834896811,
+          "f1-score": 0.9790143834001415,
+          "support": 2132.0
+        },
+        "accuracy": 0.978159509202454,
+        "macro avg": {
+          "precision": 0.9779342709417133,
+          "recall": 0.9783747690994469,
+          "f1-score": 0.9781232060259852,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9782319634681468,
+          "recall": 0.978159509202454,
+          "f1-score": 0.9781645391606395,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.0320917765745486,
+      "val_loss": 0.11052904996936644,
+      "val_accuracy": 0.9825766871165644,
+      "val_f1_weighted": 0.9825730032582375,
+      "val_f1_macro": 0.9825320474215442,
+      "val_precision_macro": 0.9827533345220337,
+      "val_recall_macro": 0.9823452205878803,
+      "hate_recall": 0.9773546062789501,
+      "hate_f1": 0.9816490049108296,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.985981308411215,
+          "recall": 0.9773546062789501,
+          "f1-score": 0.9816490049108296,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9795253606328524,
+          "recall": 0.9873358348968105,
+          "f1-score": 0.9834150899322588,
+          "support": 2132.0
+        },
+        "accuracy": 0.9825766871165644,
+        "macro avg": {
+          "precision": 0.9827533345220337,
+          "recall": 0.9823452205878803,
+          "f1-score": 0.9825320474215442,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9826036199048424,
+          "recall": 0.9825766871165644,
+          "f1-score": 0.9825730032582375,
+          "support": 4075.0
+        }
+      }
+    }
+  ],
+  "best_val_macro_f1": 0.9825320474215442,
+  "best_val_weighted_f1": 0.9825730032582375,
+  "best_epoch": 5,
+  "test_metrics": {
+    "test_accuracy": 0.9803680981595092,
+    "test_f1_macro": 0.980321168193406,
+    "test_f1_weighted": 0.9803657398697553,
+    "test_precision_macro": 0.9804425132701813,
+    "test_recall_macro": 0.9802117139604429,
+    "classification_report": {
+      "Hate": {
+        "precision": 0.9818934299017071,
+        "recall": 0.9768399382398353,
+        "f1-score": 0.9793601651186791,
+        "support": 1943.0
+      },
+      "Normal": {
+        "precision": 0.9789915966386554,
+        "recall": 0.9835834896810507,
+        "f1-score": 0.9812821712681329,
+        "support": 2132.0
+      },
+      "accuracy": 0.9803680981595092,
+      "macro avg": {
+        "precision": 0.9804425132701813,
+        "recall": 0.9802117139604429,
+        "f1-score": 0.980321168193406,
+        "support": 4075.0
+      },
+      "weighted avg": {
+        "precision": 0.9803752192227314,
+        "recall": 0.9803680981595092,
+        "f1-score": 0.9803657398697553,
+        "support": 4075.0
+      }
+    }
+  },
+  "notes": "MARBERT sweep. LR=2e-05, Batch=16. Early stopping based on validation macro F1."
+}