sft_lora_finetuner.py

"""
Fine-tune a quantized Qwen2.5:7b model using SFT + LoRA on expanded preventative health prompts.
"""

# Imports
import json
import time
from datasets import load_dataset, Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, DataCollatorForSeq2Seq
from peft import LoraConfig, get_peft_model, TaskType
import torch

# ====
# General configuration
SEED_DATA_FILE = "expanded_templates.json"  # Created locally
MODEL_NAME = "Qwen/Qwen2.5-7B"              # From Hugging Face Model Hub
OUTPUT_DIR = "./qwen_lora_adapter"
BATCH_SIZE = 4
EPOCHS = 3
LEARNING_RATE = 2e-4
MAX_LENGTH = 512
LORA_RANK = 16
LORA_ALPHA = 32
LORA_DROPOUT = 0.05

print("🔄 Starting fine-tuning pipeline for Qwen2.5 with LoRA...")
start_time = time.time()

# ===========================
# LOAD DATASET
# ===========================
print("📂 Loading dataset from:", SEED_DATA_FILE)
with open(SEED_DATA_FILE, "r", encoding="utf-8") as f:
    data = json.load(f)
print(f"✅ Loaded {len(data)} samples.")

# ====
# Convert to Hugging Face Dataset
dataset = Dataset.from_list([{
    "prompt": entry["prompt"],
    "response": entry["response"],
    "topic": entry.get("topic", "general")
} for entry in data])
print("✅ Converted to Hugging Face Dataset.")

# ====
# Concatenate prompt + response for causal LM
def tokenize_function(examples, tokenizer):

    # Takes the dictionary and formats it into a string
    texts = [
        f"### Topic: {t}\n### Instruction:\n{p}\n\n### Response:\n{r}" 
        for p, r, t in zip(examples["prompt"], examples["response"], examples["topic"])
    ]

    # Tokenize the concatenated texts
    tokenized = tokenizer(
        texts,
        max_length=MAX_LENGTH,  # Max length of tokens
        padding="max_length",
        truncation=True,
        return_tensors="pt"     # Returns input_ids, attention_maks, and labels as matrices
    )
    tokenized["labels"] = tokenized["input_ids"].clone()
    return tokenized

# ===========================
# LOAD TOKENIZER AND MODEL
# ===========================
print("🧠 Loading tokenizer and 8-bit quantized model:", MODEL_NAME)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    device_map="auto",
    torch_dtype=torch.bfloat16 
)
print("✅ Model and tokenizer loaded successfully.")

# ===========================
# CONFIGURE LoRA
# ===========================
print("⚙️ Configuring LoRA adapters...")
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=LORA_RANK,
    lora_alpha=LORA_ALPHA,
    lora_dropout=LORA_DROPOUT,
    target_modules=["q_proj", "v_proj"]  # typical for Qwen2.5
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters() 
print("✅ LoRA configuration complete.")

# ===========================
# TOKENIZE DATASET
# ===========================
print("🧩 Tokenizing dataset... (this might take a while)")
tokenized_dataset = dataset.map(lambda x: tokenize_function(x, tokenizer), batched=True)
data_collator = DataCollatorForSeq2Seq(tokenizer, pad_to_multiple_of=8, return_tensors="pt")
print("✅ Dataset tokenized and ready for training.")

# ===========================
# TRAINING ARGUMENTS
# ===========================
print("📘 Setting up training arguments...")
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    per_device_train_batch_size=BATCH_SIZE,
    gradient_accumulation_steps=2,
    learning_rate=LEARNING_RATE,
    num_train_epochs=EPOCHS,
    logging_steps=10,
    save_strategy="epoch",
    fp16=True,
    save_total_limit=3,
    report_to="none",  # disable wandb if not set up
)
print("✅ Training arguments configured.")

# ===========================
# TRAINING
# ===========================
print("🚀 Starting training...")
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    tokenizer=tokenizer,
    data_collator=data_collator
)

train_start = time.time()
trainer.train()
print(f"✅ Training completed in {(time.time() - train_start)/60:.2f} minutes.")

# ===========================
# SAVE LoRA ADAPTER ONLY
# ===========================
print("💾 Saving LoRA adapter...")
model.save_pretrained(OUTPUT_DIR)
print(f"✅ LoRA adapter saved at: {OUTPUT_DIR}")

print(f"🏁 All done! Total pipeline time: {(time.time() - start_time)/60:.2f} minutes.")