train.py

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

# --- CONFIGURATION ---
# Base model: Using a quantized Llama 3 or Mistral is recommended for consumer GPUs.
# Ensure you have access to the model on Hugging Face (might need login).
MODEL_NAME = "meta-llama/Meta-Llama-3-8B" 
DATASET_NAME = "ceperaltab/elixir-golden-dataset"
OUTPUT_DIR = "elixir-model-adapter"

def main():
    print(f"Loading dataset from {DATASET_NAME}...")
    # 1. Load Dataset
    try:
        # Load directly from HF Hub
        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" # Critical for fp16 training

    # 5. LoRA Config (Parameter Efficient Fine-Tuning)
    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"],
    )

    # 6. Formatting Function for Chat Dataset
    # Converts {"messages": [...]} into the model's expected prompt format
    def formatting_prompts_func(examples):
        output_texts = []
        for messages in examples['messages']:
            # Apply chat template (e.g., <|begin_of_text|><|start_header_id|>user...)
            # We don't tokenize yet, SFTTrainer handles it
            text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
            output_texts.append(text)
        return output_texts

    print("Starting SFTTrainer setup...")
    # 7. Trainer
    trainer = SFTTrainer(
        model=model,
        train_dataset=dataset,
        peft_config=peft_config,
        formatting_func=formatting_prompts_func,
        max_seq_length=2048,
        tokenizer=tokenizer,
        args=TrainingArguments(
            output_dir=OUTPUT_DIR,
            per_device_train_batch_size=2,
            gradient_accumulation_steps=4, # Simulate larger batch size
            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", # Change to "wandb" if desired
            push_to_hub=True,
            hub_model_id=f"ceperaltab/{OUTPUT_DIR}", # Pushes to your namespace
        ),
    )

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

if __name__ == "__main__":
    main()