train_ner.py

# /// script
# dependencies = [
#   "transformers>=4.48.0",
#   "datasets>=2.20.0",
#   "evaluate>=0.4.0",
#   "seqeval>=1.2.2",
#   "trackio",
#   "numpy<2.0",
#   "accelerate>=0.34.0",
# ]
# ///

import json
import numpy as np
import trackio
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification,
    TrainingArguments,
    Trainer,
    DataCollatorForTokenClassification,
    EarlyStoppingCallback,
)
import evaluate

MODEL_NAME   = "answerdotai/ModernBERT-base"
DATASET_NAME = "ai4privacy/pii-masking-200k"
HUB_MODEL_ID = "jefftherover/modernbert-pii-ner"
OUTPUT_DIR   = "modernbert-pii-ner"
MAX_LENGTH   = 512

# ── 1. Load full English dataset ─────────────────────────────────────────────
print("Loading dataset...")
full = load_dataset(DATASET_NAME, split="train")
en   = full.filter(lambda x: x["language"] == "en")
print(f"English rows: {len(en)}")

splits   = en.train_test_split(test_size=0.1, seed=42)
train_ds = splits["train"]
eval_ds  = splits["test"]
print(f"Train: {len(train_ds)}  Eval: {len(eval_ds)}")

# 2. Dynamic label vocabulary from data
print("Building label vocabulary...")
all_bio = set()
for ds in (train_ds, eval_ds):
    for ex in ds:
        all_bio.update(ex["mbert_bio_labels"])

label_list = (
    ["O"]
    + sorted(l for l in all_bio if l.startswith("B-"))
    + sorted(l for l in all_bio if l.startswith("I-"))
)
id2label = {i: l for i, l in enumerate(label_list)}
label2id = {l: i for i, l in id2label.items()}
print(f"Total labels: {len(label_list)}")

# 3. Tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# 4. Tokenisation + label alignment
def make_char_labels(text, raw):
    spans = json.loads(raw) if isinstance(raw, str) else raw
    cl = ["O"] * len(text)
    for span in spans:
        s, e, lbl = int(span[0]), int(span[1]), span[2]
        if lbl == "O":
            continue
        for i in range(s, min(e, len(text))):
            cl[i] = f"B-{lbl}" if i == s else f"I-{lbl}"
    return cl

def tokenize_and_align(examples):
    enc = tokenizer(
        examples["source_text"],
        truncation=True,
        max_length=MAX_LENGTH,
        return_offsets_mapping=True,
    )
    all_labels = []
    for idx in range(len(examples["source_text"])):
        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
            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
                # Word-boundary rule: new word if first token (prev_end None)
                # OR token had a leading space stripped (real_s > tok_s).
                if prev_end is None or real_s > tok_s:
                    # Word-start: assign label from char array (B-, I-, or O).
                    lbl = cl[real_s] if real_s < len(cl) else "O"
                    labels.append(label2id.get(lbl, label2id["O"]))
                else:
                    # Subword continuation (no leading space, within the same
                    # word as the previous token).  v6: if this subword falls
                    # inside an entity span give it I-<type> so the model
                    # learns to sustain entity spans across subword boundaries.
                    # Non-entity subword continuations keep -100 (ignored).
                    lbl = cl[real_s] if real_s < len(cl) else "O"
                    if lbl != "O":
                        labels.append(label2id.get(f"I-{lbl[2:]}", label2id["O"]))
                    else:
                        labels.append(-100)
                prev_end = tok_e
        all_labels.append(labels)
    enc.pop("offset_mapping")
    enc["labels"] = all_labels
    return enc

print("Tokenising datasets...")
cols      = train_ds.column_names
train_tok = train_ds.map(tokenize_and_align, batched=True, remove_columns=cols)
eval_tok  = eval_ds.map(tokenize_and_align,  batched=True, remove_columns=cols)

# 5. Metrics
seqeval = evaluate.load("seqeval")

def compute_metrics(p):
    logits, labels = p
    preds = np.argmax(logits, axis=2)
    true_preds  = [[id2label[pp] for pp, ll in zip(pr, la) if ll != -100]
                   for pr, la in zip(preds,  labels)]
    true_labels = [[id2label[ll] for pp, ll in zip(pr, la) if ll != -100]
                   for pr, la in zip(preds,  labels)]
    res = seqeval.compute(predictions=true_preds, references=true_labels)
    return {
        "precision": res["overall_precision"],
        "recall":    res["overall_recall"],
        "f1":        res["overall_f1"],
        "accuracy":  res["overall_accuracy"],
    }

# 6. Model
print("Loading model...")
model = AutoModelForTokenClassification.from_pretrained(
    MODEL_NAME,
    num_labels=len(label_list),
    id2label=id2label,
    label2id=label2id,
    ignore_mismatched_sizes=True,
)

# 7. Trackio
trackio.init(project="modernbert-pii-ner", name="modernbert-pii-ner-43k-v6")

# ── 8. Training args ─────────────────────────────────────────────────────────
# v6: subword continuation labeling fixed.
#     Previously only the first subword of each word was labeled; within-word
#     continuations got -100.  Now entity subword continuations receive I-<type>,
#     so the model learns to sustain entity spans across subword boundaries.
#     This directly targets the BOUNDARY fragmentation errors seen in v5 results
#     (e.g. "Hadley_Larson" → Had/ley/_/Larson each emitting B- instead of one span).
args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    num_train_epochs=5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=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",
    eval_strategy="steps",
    eval_steps=500,
    save_strategy="steps",
    save_steps=500,
    save_total_limit=3,
    load_best_model_at_end=True,
    metric_for_best_model="f1",
    greater_is_better=True,
    push_to_hub=True,
    hub_model_id=HUB_MODEL_ID,
    hub_strategy="every_save",
    report_to="trackio",
    run_name="modernbert-pii-ner-43k-v6",
    fp16=True,
    logging_steps=100,
    dataloader_num_workers=2,
)

# 9. Train
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=train_tok,
    eval_dataset=eval_tok,
    data_collator=DataCollatorForTokenClassification(tokenizer),
    compute_metrics=compute_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
)

print("Starting training...")
trainer.train()
trainer.push_to_hub()
print(f"Done! Model pushed to: {HUB_MODEL_ID}")