ml/train_classifier.py

"""
ClauseGuard — Fine-tune Legal-BERT on CLAUDETTE/LexGLUE unfair_tos
Multi-label classification (8 unfair clause categories)

Compatible with: Transformers 5.6.x, Datasets 4.8.x (April 2026)
"""

import os
import numpy as np
import torch
from datasets import load_dataset, Sequence, Value
from sklearn.metrics import f1_score, precision_score, recall_score
from transformers import (
    AutoConfig,
    AutoModelForSequenceClassification,
    AutoTokenizer,
    DataCollatorWithPadding,
    Trainer,
    TrainingArguments,
    EarlyStoppingCallback,
)

# ─── Config ───
MODEL_NAME = os.environ.get("BASE_MODEL", "nlpaueb/legal-bert-base-uncased")
OUTPUT_DIR = os.environ.get("OUTPUT_DIR", "./clauseguard-model")
HUB_MODEL_ID = os.environ.get("HUB_MODEL_ID", "gaurv007/clauseguard-legal-bert")
PUSH_TO_HUB = os.environ.get("PUSH_TO_HUB", "true").lower() == "true"
NUM_LABELS = 8
MAX_LENGTH = 512
LABEL_NAMES = [
    "Limitation of liability",
    "Unilateral termination",
    "Unilateral change",
    "Content removal",
    "Contract by using",
    "Choice of law",
    "Jurisdiction",
    "Arbitration",
]

print(f"ClauseGuard Model Training")
print(f"   Base model: {MODEL_NAME}")
print(f"   Output: {OUTPUT_DIR}")
print(f"   Push to Hub: {PUSH_TO_HUB} -> {HUB_MODEL_ID}")

# ─── 1. Load Dataset ───
print("Loading coastalcph/lex_glue (unfair_tos)...")
dataset = load_dataset("coastalcph/lex_glue", "unfair_tos")
print(f"   Train: {len(dataset['train'])} | Val: {len(dataset['validation'])} | Test: {len(dataset['test'])}")

# ─── 2. Load Model + Tokenizer ───
print(f"Loading {MODEL_NAME}...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

config = AutoConfig.from_pretrained(
    MODEL_NAME,
    num_labels=NUM_LABELS,
    problem_type="multi_label_classification",
    id2label={str(i): n for i, n in enumerate(LABEL_NAMES)},
    label2id={n: i for i, n in enumerate(LABEL_NAMES)},
)

model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_NAME,
    config=config,
    ignore_mismatched_sizes=True,
)
print(f"   Parameters: {sum(p.numel() for p in model.parameters()):,}")

# ─── 3. Preprocess ───
def preprocess(examples):
    tokenized = tokenizer(
        examples["text"],
        truncation=True,
        max_length=MAX_LENGTH,
        padding=False,
    )
    batch_labels = []
    for lbls in examples["labels"]:
        vec = [0.0] * NUM_LABELS
        for l in lbls:
            vec[l] = 1.0
        batch_labels.append(vec)
    tokenized["labels"] = batch_labels
    return tokenized

print("Tokenizing dataset...")
tokenized_ds = dataset.map(preprocess, batched=True, remove_columns=dataset["train"].column_names)

# Critical: cast labels to float32 for BCEWithLogitsLoss (datasets default is int64)
for split in tokenized_ds:
    tokenized_ds[split] = tokenized_ds[split].cast_column("labels", Sequence(Value("float32")))

tokenized_ds.set_format("torch")

# ─── 4. Metrics ───
def compute_metrics(eval_pred):
    logits, labels = eval_pred.predictions, eval_pred.label_ids
    probs = 1 / (1 + np.exp(-logits))
    preds = (probs > 0.5).astype(int)
    labels = labels.astype(int)
    
    micro_f1 = f1_score(labels, preds, average="micro", zero_division=0)
    macro_f1 = f1_score(labels, preds, average="macro", zero_division=0)
    micro_p = precision_score(labels, preds, average="micro", zero_division=0)
    micro_r = recall_score(labels, preds, average="micro", zero_division=0)
    
    per_class = f1_score(labels, preds, average=None, zero_division=0)
    class_metrics = {f"f1_{LABEL_NAMES[i][:15]}": float(per_class[i]) for i in range(NUM_LABELS)}
    
    return {
        "micro_f1": micro_f1,
        "macro_f1": macro_f1,
        "precision": micro_p,
        "recall": micro_r,
        **class_metrics,
    }

# ─── 5. Training ───
print("Starting training...")

training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    num_train_epochs=20,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=32,
    learning_rate=3e-5,
    weight_decay=0.01,
    warmup_ratio=0.1,
    eval_strategy="epoch",
    save_strategy="epoch",
    save_total_limit=3,
    load_best_model_at_end=True,
    metric_for_best_model="macro_f1",
    greater_is_better=True,
    fp16=torch.cuda.is_available(),
    bf16=False,
    logging_strategy="steps",
    logging_steps=25,
    logging_first_step=True,
    disable_tqdm=True,
    report_to="none",
    push_to_hub=PUSH_TO_HUB,
    hub_model_id=HUB_MODEL_ID if PUSH_TO_HUB else None,
    seed=42,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_ds["train"],
    eval_dataset=tokenized_ds["validation"],
    processing_class=tokenizer,
    data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
    compute_metrics=compute_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
)

train_result = trainer.train()
print(f"Training complete! Loss: {train_result.training_loss:.4f}")

# ─── 6. Evaluate ───
print("Evaluating on test set...")
test_results = trainer.evaluate(tokenized_ds["test"])
print(f"   Test micro-F1: {test_results.get('eval_micro_f1', 0):.4f}")
print(f"   Test macro-F1: {test_results.get('eval_macro_f1', 0):.4f}")
print(f"   Test precision: {test_results.get('eval_precision', 0):.4f}")
print(f"   Test recall: {test_results.get('eval_recall', 0):.4f}")

# ─── 7. Save ───
final_dir = f"{OUTPUT_DIR}/final"
trainer.save_model(final_dir)
tokenizer.save_pretrained(final_dir)
print(f"Model saved to {final_dir}")

if PUSH_TO_HUB:
    print(f"Pushing to Hub: {HUB_MODEL_ID}")
    trainer.push_to_hub(commit_message="ClauseGuard Legal-BERT fine-tuned on unfair_tos")
    print("Pushed successfully!")

print("Done!")