[Run 20260606_151208] Log metrics

metrics/run_20260606_151208_metrics.json

@@ -0,0 +1,250 @@
+{
+  "run_id": "20260606_151208",
+  "hyperparameters": {
+    "model_architecture": "UBC-NLP/MARBERT",
+    "task": "sequence-classification",
+    "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
+    "learning_rate": 2e-05,
+    "batch_size": 32,
+    "num_epochs": 5,
+    "max_seq_length": 128,
+    "early_stopping_patience": 2,
+    "optimizer": "AdamW",
+    "seed": 42
+  },
+  "per_epoch_results": [
+    {
+      "epoch": 1,
+      "train_loss": 0.1552482957959551,
+      "val_loss": 0.06442766166583169,
+      "val_accuracy": 0.9776687116564418,
+      "val_f1_weighted": 0.9776713498136897,
+      "val_f1_macro": 0.9776260214755199,
+      "val_precision_macro": 0.9775203320887725,
+      "val_recall_macro": 0.9777460388424701,
+      "hate_recall": 0.9794132784354092,
+      "hate_f1": 0.9766487041313832,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.973899692937564,
+          "recall": 0.9794132784354092,
+          "f1-score": 0.9766487041313832,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9811409712399811,
+          "recall": 0.976078799249531,
+          "f1-score": 0.9786033388196567,
+          "support": 2132.0
+        },
+        "accuracy": 0.9776687116564418,
+        "macro avg": {
+          "precision": 0.9775203320887725,
+          "recall": 0.9777460388424701,
+          "f1-score": 0.9776260214755199,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9776882586653561,
+          "recall": 0.9776687116564418,
+          "f1-score": 0.9776713498136897,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.06698464741152634,
+      "val_loss": 0.08235648793765904,
+      "val_accuracy": 0.979877300613497,
+      "val_f1_weighted": 0.9798606143986976,
+      "val_f1_macro": 0.97980429842375,
+      "val_precision_macro": 0.9807875544110622,
+      "val_recall_macro": 0.9791951721627354,
+      "hate_recall": 0.9644879053010808,
+      "hate_f1": 0.9785900783289817,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9931107578166402,
+          "recall": 0.9644879053010808,
+          "f1-score": 0.9785900783289817,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9684643510054844,
+          "recall": 0.9939024390243902,
+          "f1-score": 0.9810185185185185,
+          "support": 2132.0
+        },
+        "accuracy": 0.979877300613497,
+        "macro avg": {
+          "precision": 0.9807875544110622,
+          "recall": 0.9791951721627354,
+          "f1-score": 0.97980429842375,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9802159997009631,
+          "recall": 0.979877300613497,
+          "f1-score": 0.9798606143986976,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.04852369342719796,
+      "val_loss": 0.07133448649710772,
+      "val_accuracy": 0.9825766871165644,
+      "val_f1_weighted": 0.9825678862883249,
+      "val_f1_macro": 0.9825229307062366,
+      "val_precision_macro": 0.9830922659607355,
+      "val_recall_macro": 0.9821170961521564,
+      "hate_recall": 0.9722079258878024,
+      "hate_f1": 0.981553650298779,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9910807974816369,
+          "recall": 0.9722079258878024,
+          "f1-score": 0.981553650298779,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.975103734439834,
+          "recall": 0.9920262664165104,
+          "f1-score": 0.9834922111136944,
+          "support": 2132.0
+        },
+        "accuracy": 0.9825766871165644,
+        "macro avg": {
+          "precision": 0.9830922659607355,
+          "recall": 0.9821170961521564,
+          "f1-score": 0.9825229307062366,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9827217549282322,
+          "recall": 0.9825766871165644,
+          "f1-score": 0.9825678862883249,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.03323969159869843,
+      "val_loss": 0.09886962652149123,
+      "val_accuracy": 0.9791411042944785,
+      "val_f1_weighted": 0.9791271243631345,
+      "val_f1_macro": 0.9790709071924115,
+      "val_precision_macro": 0.9798448897512311,
+      "val_recall_macro": 0.9785600447654977,
+      "hate_recall": 0.9660319094184251,
+      "hate_f1": 0.9778588174003647,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9899789029535865,
+          "recall": 0.9660319094184251,
+          "f1-score": 0.9778588174003647,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9697108765488757,
+          "recall": 0.9910881801125704,
+          "f1-score": 0.9802829969844583,
+          "support": 2132.0
+        },
+        "accuracy": 0.9791411042944785,
+        "macro avg": {
+          "precision": 0.9798448897512311,
+          "recall": 0.9785600447654977,
+          "f1-score": 0.9790709071924115,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9793748704885943,
+          "recall": 0.9791411042944785,
+          "f1-score": 0.9791271243631345,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.021192557967108717,
+      "val_loss": 0.10540853567454178,
+      "val_accuracy": 0.9838036809815951,
+      "val_f1_weighted": 0.9837967629491983,
+      "val_f1_macro": 0.9837559033203549,
+      "val_precision_macro": 0.9842304403632631,
+      "val_recall_macro": 0.9834037662499433,
+      "hate_recall": 0.9747812660833762,
+      "hate_f1": 0.98287493513233,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9911041339612768,
+          "recall": 0.9747812660833762,
+          "f1-score": 0.98287493513233,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9773567467652495,
+          "recall": 0.9920262664165104,
+          "f1-score": 0.9846368715083799,
+          "support": 2132.0
+        },
+        "accuracy": 0.9838036809815951,
+        "macro avg": {
+          "precision": 0.9842304403632631,
+          "recall": 0.9834037662499433,
+          "f1-score": 0.9837559033203549,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9839116359239933,
+          "recall": 0.9838036809815951,
+          "f1-score": 0.9837967629491983,
+          "support": 4075.0
+        }
+      }
+    }
+  ],
+  "best_val_macro_f1": 0.9837559033203549,
+  "best_val_weighted_f1": 0.9837967629491983,
+  "best_epoch": 5,
+  "test_metrics": {
+    "test_accuracy": 0.9811042944785276,
+    "test_f1_macro": 0.9810417809481786,
+    "test_f1_weighted": 0.9810922726457683,
+    "test_precision_macro": 0.9817749312309503,
+    "test_recall_macro": 0.9805502795912396,
+    "classification_report": {
+      "Hate": {
+        "precision": 0.9915700737618546,
+        "recall": 0.968605249613999,
+        "f1-score": 0.9799531372038531,
+        "support": 1943.0
+      },
+      "Normal": {
+        "precision": 0.971979788700046,
+        "recall": 0.9924953095684803,
+        "f1-score": 0.9821304246925041,
+        "support": 2132.0
+      },
+      "accuracy": 0.9811042944785276,
+      "macro avg": {
+        "precision": 0.9817749312309503,
+        "recall": 0.9805502795912396,
+        "f1-score": 0.9810417809481786,
+        "support": 4075.0
+      },
+      "weighted avg": {
+        "precision": 0.981320628914793,
+        "recall": 0.9811042944785276,
+        "f1-score": 0.9810922726457683,
+        "support": 4075.0
+      }
+    }
+  },
+  "notes": "MARBERT sweep. LR=2e-05, Batch=32. Early stopping based on validation macro F1."
+}