[Run 20260606_161024] Log metrics

metrics/run_20260606_161024_metrics.json

@@ -0,0 +1,250 @@
+{
+  "run_id": "20260606_161024",
+  "hyperparameters": {
+    "model_architecture": "UBC-NLP/MARBERT",
+    "task": "sequence-classification",
+    "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
+    "learning_rate": 3e-05,
+    "batch_size": 64,
+    "num_epochs": 5,
+    "max_seq_length": 128,
+    "early_stopping_patience": 2,
+    "optimizer": "AdamW",
+    "seed": 42
+  },
+  "per_epoch_results": [
+    {
+      "epoch": 1,
+      "train_loss": 0.1463778932577299,
+      "val_loss": 0.11948525108891772,
+      "val_accuracy": 0.9739877300613496,
+      "val_f1_weighted": 0.9739560473009347,
+      "val_f1_macro": 0.9738772572783241,
+      "val_precision_macro": 0.975375976145564,
+      "val_recall_macro": 0.973064780580503,
+      "hate_recall": 0.9531652084405559,
+      "hate_f1": 0.9721784776902888,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9919657204070702,
+          "recall": 0.9531652084405559,
+          "f1-score": 0.9721784776902888,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.958786231884058,
+          "recall": 0.9929643527204502,
+          "f1-score": 0.9755760368663594,
+          "support": 2132.0
+        },
+        "accuracy": 0.9739877300613496,
+        "macro avg": {
+          "precision": 0.975375976145564,
+          "recall": 0.973064780580503,
+          "f1-score": 0.9738772572783241,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.974606537700061,
+          "recall": 0.9739877300613496,
+          "f1-score": 0.9739560473009347,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.062295065309083554,
+      "val_loss": 0.25368876694119535,
+      "val_accuracy": 0.9423312883435583,
+      "val_f1_weighted": 0.9419861159442119,
+      "val_f1_macro": 0.9417061263155575,
+      "val_precision_macro": 0.950236143565397,
+      "val_recall_macro": 0.9395493178475869,
+      "hate_recall": 0.8795676788471436,
+      "hate_f1": 0.9356693128935122,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9994152046783625,
+          "recall": 0.8795676788471436,
+          "f1-score": 0.9356693128935122,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9010570824524313,
+          "recall": 0.9995309568480301,
+          "f1-score": 0.9477429397376028,
+          "support": 2132.0
+        },
+        "accuracy": 0.9423312883435583,
+        "macro avg": {
+          "precision": 0.950236143565397,
+          "recall": 0.9395493178475869,
+          "f1-score": 0.9417061263155575,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9479552006082557,
+          "recall": 0.9423312883435583,
+          "f1-score": 0.9419861159442119,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.042429992100227805,
+      "val_loss": 0.08841686587220465,
+      "val_accuracy": 0.9815950920245399,
+      "val_f1_weighted": 0.9815846068456822,
+      "val_f1_macro": 0.9815362727285084,
+      "val_precision_macro": 0.9821854148248047,
+      "val_recall_macro": 0.9810877600739268,
+      "hate_recall": 0.9701492537313433,
+      "hate_f1": 0.9804941482444733,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.991062039957939,
+          "recall": 0.9701492537313433,
+          "f1-score": 0.9804941482444733,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9733087896916705,
+          "recall": 0.9920262664165104,
+          "f1-score": 0.9825783972125436,
+          "support": 2132.0
+        },
+        "accuracy": 0.9815950920245399,
+        "macro avg": {
+          "precision": 0.9821854148248047,
+          "recall": 0.9810877600739268,
+          "f1-score": 0.9815362727285084,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9817737136836604,
+          "recall": 0.9815950920245399,
+          "f1-score": 0.9815846068456822,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.02510130808260151,
+      "val_loss": 0.0717304450595293,
+      "val_accuracy": 0.9852760736196319,
+      "val_f1_weighted": 0.9852723575092364,
+      "val_f1_macro": 0.9852372402659978,
+      "val_precision_macro": 0.9855061116942819,
+      "val_recall_macro": 0.9850162076980047,
+      "hate_recall": 0.9794132784354092,
+      "hate_f1": 0.9844800827728919,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9895995839833593,
+          "recall": 0.9794132784354092,
+          "f1-score": 0.9844800827728919,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9814126394052045,
+          "recall": 0.9906191369606003,
+          "f1-score": 0.9859943977591037,
+          "support": 2132.0
+        },
+        "accuracy": 0.9852760736196319,
+        "macro avg": {
+          "precision": 0.9855061116942819,
+          "recall": 0.9850162076980047,
+          "f1-score": 0.9852372402659978,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9853162549427149,
+          "recall": 0.9852760736196319,
+          "f1-score": 0.9852723575092364,
+          "support": 4075.0
+        }
+      }
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.020272867037444343,
+      "val_loss": 0.10328626508123762,
+      "val_accuracy": 0.9850306748466258,
+      "val_f1_weighted": 0.9850294351740828,
+      "val_f1_macro": 0.9849959640154234,
+      "val_precision_macro": 0.9850792204204483,
+      "val_recall_macro": 0.9849185607834542,
+      "hate_recall": 0.9825012866700978,
+      "hate_f1": 0.9842742974993555,
+      "classification_report": {
+        "Hate": {
+          "precision": 0.9860537190082644,
+          "recall": 0.9825012866700978,
+          "f1-score": 0.9842742974993555,
+          "support": 1943.0
+        },
+        "Normal": {
+          "precision": 0.9841047218326321,
+          "recall": 0.9873358348968105,
+          "f1-score": 0.9857176305314914,
+          "support": 2132.0
+        },
+        "accuracy": 0.9850306748466258,
+        "macro avg": {
+          "precision": 0.9850792204204483,
+          "recall": 0.9849185607834542,
+          "f1-score": 0.9849959640154234,
+          "support": 4075.0
+        },
+        "weighted avg": {
+          "precision": 0.9850340228172343,
+          "recall": 0.9850306748466258,
+          "f1-score": 0.9850294351740828,
+          "support": 4075.0
+        }
+      }
+    }
+  ],
+  "best_val_macro_f1": 0.9852372402659978,
+  "best_val_weighted_f1": 0.9852723575092364,
+  "best_epoch": 4,
+  "test_metrics": {
+    "test_accuracy": 0.9850306748466258,
+    "test_f1_macro": 0.9849884977742274,
+    "test_f1_weighted": 0.9850254027125759,
+    "test_precision_macro": 0.9853737715887565,
+    "test_recall_macro": 0.9846904363477302,
+    "classification_report": {
+      "Hate": {
+        "precision": 0.9911273486430062,
+        "recall": 0.9773546062789501,
+        "f1-score": 0.9841927960611557,
+        "support": 1943.0
+      },
+      "Normal": {
+        "precision": 0.9796201945345067,
+        "recall": 0.9920262664165104,
+        "f1-score": 0.985784199487299,
+        "support": 2132.0
+      },
+      "accuracy": 0.9850306748466258,
+      "macro avg": {
+        "precision": 0.9853737715887565,
+        "recall": 0.9846904363477302,
+        "f1-score": 0.9849884977742274,
+        "support": 4075.0
+      },
+      "weighted avg": {
+        "precision": 0.9851069185670992,
+        "recall": 0.9850306748466258,
+        "f1-score": 0.9850254027125759,
+        "support": 4075.0
+      }
+    }
+  },
+  "notes": "MARBERT sweep. LR=3e-05, Batch=64. Early stopping based on validation macro F1."
+}