train.py

import torch
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig
import trackio

# Configuration
MODEL_ID = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
DATASET_ID = "iamtarun/code_instructions_120k_alpaca"
OUTPUT_DIR = "./qwen-coder-multilingual-sft"
HUB_MODEL_ID = "moos124/qwen-coder-multilingual-sft"

def preprocess_function(example):
    # Convert Alpaca format to ChatML format
    user_content = example["instruction"]
    if example.get("input"):
        user_content += f"\n\nInput: {example['input']}"
    
    return {
        "messages": [
            {"role": "user", "content": user_content},
            {"role": "assistant", "content": example["output"]}
        ]
    }

def main():
    # 1. Load Dataset
    dataset = load_dataset(DATASET_ID, split="train")
    dataset = dataset.map(preprocess_function, remove_columns=dataset.column_names)

    # 2. Tokenizer & Model
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    tokenizer.pad_token = tokenizer.eos_token
    
    # 3. PEFT Config (LoRA)
    peft_config = LoraConfig(
        r=16,
        lora_alpha=32,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM",
    )

    # 4. SFTConfig
    sft_config = SFTConfig(
        output_dir=OUTPUT_DIR,
        max_seq_length=2048,
        dataset_text_field="messages",
        packing=False,
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        learning_rate=2e-4,
        num_train_epochs=1,
        weight_decay=0.01,
        lr_scheduler_type="cosine",
        warmup_steps=100,
        logging_steps=10,
        logging_first_step=True,
        disable_tqdm=True,
        logging_strategy="steps",
        bf16=True,
        gradient_checkpointing=True,
        push_to_hub=True,
        hub_model_id=HUB_MODEL_ID,
        save_strategy="steps",
        save_steps=500,
        report_to="trackio",
    )

    # 5. Trainer
    trainer = SFTTrainer(
        model=MODEL_ID,
        train_dataset=dataset,
        args=sft_config,
        peft_config=peft_config,
        processing_class=tokenizer,
    )

    # 6. Train
    trainer.train()
    
    # 7. Save & Push
    trainer.save_model(OUTPUT_DIR)
    trainer.push_to_hub()

if __name__ == "__main__":
    main()