train.py

"""Simple training script for a text toxicity classifier.

Usage examples:
  - Train from a CSV: python train.py --dataset_csv data/toxic_train.csv --text_col text --label_col label --output_dir ./outputs
  - Push to Hub: python train.py --dataset_csv data/toxic_train.csv --output_dir ./outputs --push_to_hub --hub_model_id your-username/toxic-detector

Expect CSV with columns: text, label (0/1) for single-label classification. For multi-label adjust the preprocessing.
"""
import argparse
from pathlib import Path
from datasets import load_dataset, Dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    TrainingArguments,
    Trainer,
    DataCollatorWithPadding,
)
import numpy as np
import evaluate


def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--dataset_csv", type=str, default=None, help="Path to CSV dataset with text and label columns")
    p.add_argument("--text_col", type=str, default="text")
    p.add_argument("--label_col", type=str, default="label")
    p.add_argument("--model_name_or_path", type=str, default="distilbert-base-uncased")
    p.add_argument("--output_dir", type=str, default="./model_output")
    p.add_argument("--push_to_hub", action="store_true")
    p.add_argument("--hub_model_id", type=str, default=None)
    p.add_argument("--num_train_epochs", type=int, default=1)
    p.add_argument("--per_device_train_batch_size", type=int, default=16)
    return p.parse_args()


def main():
    args = parse_args()

    if args.dataset_csv:
        ds = load_dataset("csv", data_files={"train": args.dataset_csv})
        # if no validation split, take 10% for val
        ds = ds["train"].train_test_split(test_size=0.1)
        dataset = ds
    else:
        # small built-in fallback: use a tiny subset of imdb for demo (binary sentiment)
        dataset = load_dataset("imdb", split={"train": "train[:2000]","test": "test[:500]"})
        dataset = dataset

    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)

    def preprocess_function(examples):
        texts = examples[args.text_col] if args.dataset_csv else examples["text"]
        return tokenizer(texts, truncation=True)

    if args.dataset_csv:
        tokenized = dataset.map(preprocess_function, batched=True)
    else:
        # imdb default has 'text' and 'label'
        tokenized = dataset.map(lambda x: tokenizer(x['text'], truncation=True), batched=True)

    labels = tokenized["train"].features[args.label_col] if args.dataset_csv else None

    num_labels = 2
    model = AutoModelForSequenceClassification.from_pretrained(args.model_name_or_path, num_labels=num_labels)

    metric_acc = evaluate.load("accuracy")

    def compute_metrics(eval_pred):
        logits, labels = eval_pred
        preds = np.argmax(logits, axis=-1)
        return metric_acc.compute(predictions=preds, references=labels)

    data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

    training_args = TrainingArguments(
        output_dir=args.output_dir,
        evaluation_strategy="epoch",
        num_train_epochs=args.num_train_epochs,
        per_device_train_batch_size=args.per_device_train_batch_size,
        save_total_limit=2,
        push_to_hub=args.push_to_hub,
        hub_model_id=args.hub_model_id,
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized["train"],
        eval_dataset=tokenized.get("test", None),
        tokenizer=tokenizer,
        data_collator=data_collator,
        compute_metrics=compute_metrics,
    )

    trainer.train()

    trainer.save_model()
    if args.push_to_hub and args.hub_model_id:
        trainer.push_to_hub()


if __name__ == "__main__":
    main()