Upload train_ner.py with huggingface_hub

train_ner.py

@@ -20,6 +20,7 @@ from transformers import (
     TrainingArguments,
     Trainer,
     DataCollatorForTokenClassification,
+    EarlyStoppingCallback,
 )
 import evaluate
 
@@ -132,19 +133,20 @@ model = AutoModelForTokenClassification.from_pretrained(
 )
 
 # 7. Trackio
-trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v2")
+trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v3")
 
 # ── 8. Training args ─────────────────────────────────────────────────────────
 args = TrainingArguments(
     output_dir=OUTPUT_DIR,
-    num_train_epochs=5,
+    num_train_epochs=3,                    # v2 peaked at ~epoch 2.45; cap at 3
     per_device_train_batch_size=16,
     per_device_eval_batch_size=32,
-    gradient_accumulation_steps=2,        # effective batch = 32
+    gradient_accumulation_steps=2,         # effective batch = 32
     learning_rate=5e-5,
     weight_decay=0.01,
     warmup_ratio=0.2,
     lr_scheduler_type="cosine_with_restarts",
+    label_smoothing_factor=0.1,            # reduces overconfidence / overfitting
     eval_strategy="steps",
     eval_steps=500,
     save_strategy="steps",
@@ -157,7 +159,7 @@ args = TrainingArguments(
     hub_model_id=HUB_MODEL_ID,
     hub_strategy="every_save",
     report_to="trackio",
-    run_name="modernbert-pii-ner-43k-v2",
+    run_name="modernbert-pii-ner-43k-v3",
     fp16=True,
     logging_steps=100,
     dataloader_num_workers=2,
@@ -171,6 +173,7 @@ trainer = Trainer(
     eval_dataset=eval_tok,
     data_collator=DataCollatorForTokenClassification(tokenizer),
     compute_metrics=compute_metrics,
+    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
 )
 
 print("Starting training...")