train.py

import os
import torch
from datasets import load_dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig

# --- CONFIGURATION ---
MODEL_NAME = "Qwen/Qwen2.5-Coder-7B-Instruct" 
DATASET_NAME = "ceperaltab/elixir-golden-dataset"
OUTPUT_DIR = "elixir-model-qwen"

def main():
    print(f"Loading dataset from {DATASET_NAME}...")
    # 1. Load Dataset
    try:
        dataset = load_dataset(DATASET_NAME, split="train")
    except Exception as e:
        print(f"Error loading dataset: {e}")
        return

    # 2. Quantization Config (4-bit for memory efficiency)
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.float16,
    )

    print(f"Loading base model: {MODEL_NAME}...")
    # 3. Load Model
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=True
    )
    
    # 4. Load Tokenizer
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    # 5. LoRA Config
    peft_config = LoraConfig(
        lora_alpha=16,
        lora_dropout=0.1,
        r=64,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    )

    print("Starting SFTTrainer setup...")
    
    # 6. SFTConfig (new TRL API - all config goes here)
    sft_config = SFTConfig(
        output_dir=OUTPUT_DIR,
        max_seq_length=2048,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        learning_rate=2e-4,
        logging_steps=10,
        num_train_epochs=1,
        optim="paged_adamw_32bit",
        fp16=True,
        group_by_length=True,
        save_strategy="epoch",
        report_to="none",
        push_to_hub=True,
        hub_model_id=f"ceperaltab/{OUTPUT_DIR}",
    )

    # 7. Trainer - new API: use processing_class instead of tokenizer
    # The trainer automatically handles conversational datasets with "messages" field
    trainer = SFTTrainer(
        model=model,
        args=sft_config,
        train_dataset=dataset,
        processing_class=tokenizer,  # New API: processing_class replaces tokenizer
        peft_config=peft_config,
    )

    print("Starting training...")
    trainer.train()
    
    print(f"Saving model to {OUTPUT_DIR}...")
    trainer.save_model(OUTPUT_DIR)
    print("Done!")

if __name__ == "__main__":
    main()