v4: fix space-offset alignment bug in tokenize_and_align; revert to v2 hyperparams (5 epochs, no label smoothing)

train_ner.py

@@ -81,15 +81,22 @@ def tokenize_and_align(examples):
     )
     all_labels = []
     for idx in range(len(examples["source_text"])):
-        cl      = make_char_labels(examples["source_text"][idx],
-                                   examples["span_labels"][idx])
+        text    = examples["source_text"][idx]
+        cl      = make_char_labels(text, examples["span_labels"][idx])
         offsets = enc["offset_mapping"][idx]
         labels, prev_end = [], None
         for tok_s, tok_e in offsets:
             if tok_s == tok_e:
                 labels.append(-100); prev_end = None
             elif prev_end is None or tok_s > prev_end:
-                lbl = cl[tok_s] if tok_s < len(cl) else "O"
+                # ModernBERT tokenizer includes the preceding space in the
+                # token's offset (e.g. " Grey" → offset starts at the space).
+                # Advance past any leading spaces so we look up the actual
+                # first character of the word in the char-label array.
+                real_s = tok_s
+                while real_s < tok_e and text[real_s] == ' ':
+                    real_s += 1
+                lbl = cl[real_s] if real_s < len(cl) else "O"
                 labels.append(label2id.get(lbl, label2id["O"]))
                 prev_end = tok_e
             else:
@@ -133,12 +140,15 @@ model = AutoModelForTokenClassification.from_pretrained(
 )
 
 # 7. Trackio
-trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v3")
+trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v4")
 
 # ── 8. Training args ─────────────────────────────────────────────────────────
+# v4: alignment bug fixed (space-stripped char-label lookup).
+#     Hyperparams revert to v2 best (5 epochs, no label smoothing — v3's
+#     label_smoothing_factor=0.1 hurt precision more than it helped recall).
 args = TrainingArguments(
     output_dir=OUTPUT_DIR,
-    num_train_epochs=3,                    # v2 peaked at ~epoch 2.45; cap at 3
+    num_train_epochs=5,                    # v2 peaked at epoch 4.09; give full budget
     per_device_train_batch_size=16,
     per_device_eval_batch_size=32,
     gradient_accumulation_steps=2,         # effective batch = 32
@@ -146,7 +156,7 @@ args = TrainingArguments(
     weight_decay=0.01,
     warmup_ratio=0.2,
     lr_scheduler_type="cosine_with_restarts",
-    label_smoothing_factor=0.1,            # reduces overconfidence / overfitting
+    # label_smoothing_factor removed — hurt v3 precision by -1.93%
     eval_strategy="steps",
     eval_steps=500,
     save_strategy="steps",
@@ -159,7 +169,7 @@ args = TrainingArguments(
     hub_model_id=HUB_MODEL_ID,
     hub_strategy="every_save",
     report_to="trackio",
-    run_name="modernbert-pii-ner-43k-v3",
+    run_name="modernbert-pii-ner-43k-v4",
     fp16=True,
     logging_steps=100,
     dataloader_num_workers=2,