train.py

import torch
import json
import os
from datasets import Dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
from huggingface_hub import login

# ✅ Authenticate with Hugging Face
HF_TOKEN = os.getenv("HF_TOKEN")

if not HF_TOKEN:
    raise ValueError("Hugging Face token not found. Add it in 'Secrets'.")

login(token=HF_TOKEN)

# ✅ Load Extracted Data
dataset_path = "medical_dataset.json"
if not os.path.exists(dataset_path):
    raise FileNotFoundError(f"Dataset file '{dataset_path}' not found!")

with open(dataset_path, "r", encoding="utf-8") as f:
    data = json.load(f)

if not isinstance(data, list):
    raise ValueError("Dataset should be a list of dictionaries.")

dataset = Dataset.from_list(data)

# ✅ Load Tokenizer
model_name = "tiiuae/falcon-rw-1b"
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# ✅ Tokenize Data (Fixed)
def preprocess_function(examples):
    prompt = examples.get("prompt", "")
    response = examples.get("response", "")
    inputs = f"Medical Q&A: {prompt} {response}"
    
    model_inputs = tokenizer(inputs, padding="max_length", truncation=True, max_length=512)

    # ✅ Ensure labels have the same length
    model_inputs["labels"] = model_inputs["input_ids"]

    return {key: [val] for key, val in model_inputs.items()}  # ✅ Wrap values in lists

# ✅ Apply tokenization
tokenized_dataset = dataset.map(preprocess_function, batched=True, remove_columns=dataset.column_names)


# ✅ Load Model with LoRA (Optimized for Falcon)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,  # ✅ Save memory
    device_map="auto"  # ✅ Auto-assign to CPU/GPU
)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["query_key_value"],  # ✅ Correct target module for Falcon
    lora_dropout=0.05,
    bias="none"
)

model = get_peft_model(model, lora_config)

# ✅ Define Training Arguments
training_args = TrainingArguments(
    output_dir="./medical_falcon",
    per_device_train_batch_size=1,
    num_train_epochs=3,  # ✅ Adjust epochs as needed
    logging_dir="./logs",
    save_steps=100,
    evaluation_strategy="no",
    save_total_limit=2,
    fp16=True  # ✅ Enable mixed precision training
)

# ✅ Train Model
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset
)

trainer.train()

# ✅ Save Model
model_path = "fine_tuned_medical_falcon"
trainer.save_model(model_path)
tokenizer.save_pretrained(model_path)

print(f"✅ Model fine-tuned and saved at: {model_path}")