train.py

import os
import torch
from datasets import load_dataset, Dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
)
from peft import LoraConfig
from trl.trainer.sft_trainer import SFTTrainer
from trl.trainer.sft_config import SFTConfig
import argparse
import pandas as pd

# Define tokenizer globally for the mapping function
tokenizer = None


def format_instruction(sample):
    # Standard format for SmolLM2-Instruct
    label_str = "Phishing" if sample["phishing"] == 1 else "Safe"

    messages = [
        {
            "role": "user",
            "content": f"Classify the following email text as either 'Safe' or 'Phishing'. Respond with only one word: 'Safe' or 'Phishing'.\n\nEmail text: {sample['text']}\n\nClassification:",
        },
        {"role": "assistant", "content": label_str},
    ]
    # tokenizer is now accessible globally
    return (
        {"text": tokenizer.apply_chat_template(messages, tokenize=False)}
        if tokenizer
        else {"text": ""}
    )


def main(args):
    global tokenizer
    device = (
        "cuda"
        if torch.cuda.is_available()
        else "mps"
        if torch.backends.mps.is_available()
        else "cpu"
    )
    print(f"Using device: {device}")

    model_id = args.model_id
    print(f"Loading tokenizer and model: {model_id}")

    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.pad_token = tokenizer.eos_token

    # Load Model
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        device_map=device if device != "mps" else None,
    )
    if device == "mps":
        model.to("mps")  # type: ignore

    # LoRA Configuration
    peft_config = LoraConfig(
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        lora_dropout=args.lora_dropout,
        target_modules=[
            "q_proj",
            "k_proj",
            "v_proj",
            "o_proj",
            "gate_proj",
            "up_proj",
            "down_proj",
        ],
        bias="none",
        task_type="CAUSAL_LM",
    )

    # Load Data
    print(f"Loading data from {args.dataset_name}...")
    if os.path.exists(args.dataset_name):
        train_df = pd.read_csv(os.path.join(args.dataset_name, "train.csv"))
        val_df = pd.read_csv(os.path.join(args.dataset_name, "val.csv"))
        if args.quick_test:
            train_df = train_df.head(100)
            val_df = val_df.head(20)
        train_dataset = Dataset.from_pandas(train_df)
        val_dataset = Dataset.from_pandas(val_df)
    else:
        dataset = load_dataset(args.dataset_name)
        train_dataset = dataset["train"]
        val_dataset = dataset["validation"] if "validation" in dataset else None

    # Apply formatting
    print("Formatting datasets...")
    train_dataset = train_dataset.map(format_instruction)
    if val_dataset:
        val_dataset = val_dataset.map(format_instruction)

    # Use SFTConfig for modern TRL
    sft_config = SFTConfig(
        output_dir=args.output_dir,
        per_device_train_batch_size=args.batch_size,
        gradient_accumulation_steps=args.grad_accum,
        learning_rate=args.lr,
        logging_steps=10,
        num_train_epochs=args.epochs,
        max_steps=args.max_steps,
        eval_strategy="steps" if val_dataset else "no",
        eval_steps=100,
        save_strategy="steps",
        save_steps=100,
        lr_scheduler_type="cosine",
        warmup_ratio=0.1,
        bf16=torch.cuda.is_available(),
        push_to_hub=args.push_to_hub,
        report_to="tensorboard" if not args.no_report else "none",
        remove_unused_columns=False,
        dataset_text_field="text",
        max_length=args.max_seq_length,
    )

    # Standard HF SFTTrainer
    trainer = SFTTrainer(
        model=model,
        train_dataset=train_dataset,
        eval_dataset=val_dataset,
        peft_config=peft_config,
        processing_class=tokenizer,
        args=sft_config,
    )

    print("Starting training...")
    trainer.train()

    print(f"Saving model to {args.output_dir}")
    trainer.save_model(args.output_dir)
    if args.push_to_hub:
        trainer.push_to_hub()


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_id", type=str, default="HuggingFaceTB/SmolLM2-135M-Instruct"
    )
    parser.add_argument("--dataset_name", type=str, default="data/")
    parser.add_argument("--output_dir", type=str, default="models/smollm2-phish-sft")
    parser.add_argument("--batch_size", type=int, default=4)
    parser.add_argument("--grad_accum", type=int, default=4)
    parser.add_argument("--lr", type=float, default=2e-4)
    parser.add_argument("--epochs", type=int, default=1)
    parser.add_argument("--max_steps", type=int, default=-1)
    parser.add_argument("--max_seq_length", type=int, default=512)
    parser.add_argument("--lora_r", type=int, default=16)
    parser.add_argument("--lora_alpha", type=int, default=32)
    parser.add_argument("--lora_dropout", type=float, default=0.05)
    parser.add_argument("--quick_test", action="store_true")
    parser.add_argument("--push_to_hub", action="store_true")
    parser.add_argument("--no_report", action="store_true")
    args = parser.parse_args()
    main(args)