Upload train_ner.py with huggingface_hub

train_ner.py

@@ -88,19 +88,26 @@ def tokenize_and_align(examples):
         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:
-                # 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.
+            else:
+                # ModernBERT (RoBERTa-style) tokenizer absorbs the preceding
+                # space into the next token's offset (e.g. " Grey" has ts at
+                # the space, not at 'G').  Strip leading spaces to find the
+                # true first character of the word.
                 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"]))
+                # Word-boundary rule: new word if this is the very first token
+                # (prev_end is None) OR the token had a leading space stripped
+                # (real_s > tok_s).  The old check (tok_s > prev_end) was wrong
+                # for RoBERTa tokenizers: spaces are inside the token so there
+                # is never a gap between consecutive tokens, meaning only the
+                # very first token per sentence was ever labeled.
+                if prev_end is None or real_s > tok_s:
+                    lbl = cl[real_s] if real_s < len(cl) else "O"
+                    labels.append(label2id.get(lbl, label2id["O"]))
+                else:
+                    labels.append(-100)
                 prev_end = tok_e
-            else:
-                labels.append(-100); prev_end = tok_e
         all_labels.append(labels)
     enc.pop("offset_mapping")
     enc["labels"] = all_labels
@@ -140,15 +147,18 @@ model = AutoModelForTokenClassification.from_pretrained(
 )
 
 # 7. Trackio
-trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v4")
+trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v5")
 
 # ── 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).
+# v5: word-boundary detection fixed (real_s > tok_s instead of tok_s > prev_end).
+#     The old check never fired for RoBERTa-style tokenizers where spaces are
+#     absorbed into the next token's offset — only the first token per sentence
+#     was ever labeled (effectively 1 label per example).  With the fix every
+#     word gets labeled, giving the model a proper training signal for in-sentence
+#     and multi-word entities.
 args = TrainingArguments(
     output_dir=OUTPUT_DIR,
-    num_train_epochs=5,                    # v2 peaked at epoch 4.09; give full budget
+    num_train_epochs=5,
     per_device_train_batch_size=16,
     per_device_eval_batch_size=32,
     gradient_accumulation_steps=2,         # effective batch = 32
@@ -156,7 +166,6 @@ args = TrainingArguments(
     weight_decay=0.01,
     warmup_ratio=0.2,
     lr_scheduler_type="cosine_with_restarts",
-    # label_smoothing_factor removed — hurt v3 precision by -1.93%
     eval_strategy="steps",
     eval_steps=500,
     save_strategy="steps",
@@ -169,7 +178,7 @@ args = TrainingArguments(
     hub_model_id=HUB_MODEL_ID,
     hub_strategy="every_save",
     report_to="trackio",
-    run_name="modernbert-pii-ner-43k-v4",
+    run_name="modernbert-pii-ner-43k-v5",
     fp16=True,
     logging_steps=100,
     dataloader_num_workers=2,