Files
LLM_Engineering_OLD/week6/community-contributions/salah/smart_fine_tuner.py

269 lines
9.8 KiB
Python

import sys
import os
sys.path.append("../..")
import json
import pickle
import pandas as pd
import numpy as np
from openai import OpenAI
from dotenv import load_dotenv
from huggingface_hub import login
from smart_pricer import SmartPricer, ConfidenceAwareTester
import re
from typing import List, Dict, Tuple
import time
load_dotenv(override=True)
os.environ['OPENAI_API_KEY'] = os.getenv('OPENAI_API_KEY')
os.environ['HF_TOKEN'] = os.getenv('HF_TOKEN')
hf_token = os.environ['HF_TOKEN']
login(hf_token, add_to_git_credential=True)
class SmartFineTuner:
def __init__(self, openai_api_key: str = None):
self.client = OpenAI(api_key=openai_api_key or os.getenv('OPENAI_API_KEY'))
self.fine_tuned_model_id = None
self.training_templates = [
{
"system": "You are a product pricing expert. Respond only with the price, no explanation.",
"user": "Estimate the price of this product:\n\n{description}\n\nPrice: $",
"weight": 0.4
},
{
"system": "You are a retail pricing expert. Consider market positioning and consumer value.",
"user": "What would this product sell for in the market?\n\n{description}\n\nMarket price: $",
"weight": 0.3
},
{
"system": "You analyze product features to determine fair pricing.",
"user": "Based on the features and quality described, estimate the price:\n\n{description}\n\nEstimated price: $",
"weight": 0.3
}
]
def prepare_enhanced_training_data(self, train_items: List, template_mix: bool = True) -> List[Dict]:
training_data = []
for item in train_items:
description = self._get_clean_description(item)
if len(description.strip()) < 20:
continue
if hasattr(item, 'price'):
price = item.price
else:
price = item.get('price', 0)
if price <= 0:
continue
templates_to_use = self.training_templates if template_mix else [self.training_templates[0]]
for template in templates_to_use:
if template_mix and np.random.random() > template['weight']:
continue
user_prompt = template['user'].format(description=description)
messages = [
{"role": "system", "content": template['system']},
{"role": "user", "content": user_prompt},
{"role": "assistant", "content": f"{price:.2f}"}
]
training_data.append({"messages": messages})
return training_data
def _get_clean_description(self, item) -> str:
if hasattr(item, 'test_prompt'):
prompt = item.test_prompt()
clean = prompt.replace(" to the nearest dollar", "")
clean = clean.replace("\n\nPrice is $", "")
clean = re.sub(r'\$\d+\.?\d*', '', clean)
clean = re.sub(r'\d+\.?\d*\s*dollars?', '', clean)
return clean.strip()
else:
parts = []
if 'title' in item and item['title']:
parts.append(f"Title: {item['title']}")
if 'description' in item and item['description']:
parts.append(f"Description: {item['description']}")
if 'features' in item and item['features']:
parts.append(f"Features: {item['features']}")
return '\n'.join(parts)
def create_training_files(self, train_items: List, val_items: List,
enhanced: bool = True) -> Tuple[str, str]:
train_data = self.prepare_enhanced_training_data(train_items, template_mix=enhanced)
val_data = self.prepare_enhanced_training_data(val_items, template_mix=False)
print(f"Prepared {len(train_data)} training examples")
print(f"Prepared {len(val_data)} validation examples")
train_file = "smart_pricer_train.jsonl"
val_file = "smart_pricer_validation.jsonl"
with open(train_file, 'w') as f:
for example in train_data:
f.write(json.dumps(example) + '\n')
with open(val_file, 'w') as f:
for example in val_data:
f.write(json.dumps(example) + '\n')
return train_file, val_file
def start_fine_tuning(self, train_file: str, val_file: str,
model: str = "gpt-4o-mini-2024-07-18",
epochs: int = 1) -> str:
print(f"Starting fine-tuning with enhanced training data...")
with open(train_file, 'rb') as f:
train_file_obj = self.client.files.create(file=f, purpose="fine-tune")
with open(val_file, 'rb') as f:
val_file_obj = self.client.files.create(file=f, purpose="fine-tune")
print(f"Uploaded training file: {train_file_obj.id}")
print(f"Uploaded validation file: {val_file_obj.id}")
job = self.client.fine_tuning.jobs.create(
training_file=train_file_obj.id,
validation_file=val_file_obj.id,
model=model,
hyperparameters={"n_epochs": epochs},
suffix="smart_pricer"
)
self.fine_tuned_model_id = job.id
print(f"Fine-tuning job created: {job.id}")
return job.id
def check_job_status(self, job_id: str) -> Dict:
job = self.client.fine_tuning.jobs.retrieve(job_id)
return {
'status': job.status,
'model': job.fine_tuned_model,
'created_at': job.created_at,
'finished_at': job.finished_at
}
def evaluate_fine_tuned_model(self, test_data: List, job_id: str) -> Dict:
job_info = self.check_job_status(job_id)
if job_info['status'] != 'succeeded':
print(f"Job not completed yet. Status: {job_info['status']}")
return {}
fine_tuned_model = job_info['model']
print(f"Evaluating fine-tuned model: {fine_tuned_model}")
pricer = SmartPricer(fine_tuned_model=fine_tuned_model)
tester = ConfidenceAwareTester(
pricer,
test_data[:100],
title=f"Fine-tuned Smart Pricer ({fine_tuned_model})",
size=100
)
results = tester.run_enhanced_test()
if results:
avg_error = np.mean([r['error'] for r in results])
avg_confidence = np.mean([r['confidence'] for r in results])
high_conf_results = [r for r in results if r['confidence'] > 0.7]
high_conf_error = np.mean([r['error'] for r in high_conf_results]) if high_conf_results else float('inf')
summary = {
'model_id': fine_tuned_model,
'total_predictions': len(results),
'average_error': avg_error,
'average_confidence': avg_confidence,
'high_confidence_count': len(high_conf_results),
'high_confidence_error': high_conf_error,
'job_id': job_id
}
print(f"\nEVALUATION SUMMARY:")
print(f"Average Error: ${avg_error:.2f}")
print(f"Average Confidence: {avg_confidence:.2f}")
print(f"High Confidence Predictions: {len(high_conf_results)}")
print(f"High Confidence Error: ${high_conf_error:.2f}")
return summary
return {}
def quick_fine_tune_demo(train_size: int = 200, val_size: int = 50):
print("Smart Pricer Fine-Tuning Demo")
print("=" * 50)
try:
with open('train.pkl', 'rb') as file:
train_data = pickle.load(file)
with open('test.pkl', 'rb') as file:
test_data = pickle.load(file)
print(f"Loaded training data: {len(train_data)} items")
print(f"Loaded test data: {len(test_data)} items")
except FileNotFoundError:
print("Training data not found. Make sure train.pkl and test.pkl are in current directory.")
return
train_items = train_data[:train_size]
val_items = train_data[train_size:train_size + val_size]
print(f"Using {len(train_items)} training items, {len(val_items)} validation items")
fine_tuner = SmartFineTuner()
train_file, val_file = fine_tuner.create_training_files(
train_items, val_items, enhanced=True
)
print(f"Created training files: {train_file}, {val_file}")
print(f"\nTo start fine-tuning, uncomment the following lines:")
print(f"job_id = fine_tuner.start_fine_tuning('{train_file}', '{val_file}')")
print(f"# Wait for job to complete...")
print(f"# results = fine_tuner.evaluate_fine_tuned_model(test_data, job_id)")
print(f"\nDemo with base model (no fine-tuning):")
pricer = SmartPricer()
tester = ConfidenceAwareTester(pricer, test_data[:25], size=25)
tester.run_enhanced_test()
def main():
import argparse
parser = argparse.ArgumentParser(description='Smart Pricer Fine-Tuning')
parser.add_argument('--demo', action='store_true', help='Run demo mode')
parser.add_argument('--train-size', type=int, default=200, help='Training set size')
parser.add_argument('--val-size', type=int, default=50, help='Validation set size')
parser.add_argument('--evaluate', type=str, help='Evaluate existing model by job ID')
args = parser.parse_args()
if args.demo:
quick_fine_tune_demo(args.train_size, args.val_size)
elif args.evaluate:
try:
with open('test.pkl', 'rb') as file:
test_data = pickle.load(file)
fine_tuner = SmartFineTuner()
fine_tuner.evaluate_fine_tuned_model(test_data, args.evaluate)
except FileNotFoundError:
print("Test data not found. Make sure test.pkl is in current directory.")
else:
print("Use --demo to run demo or --evaluate <job_id> to evaluate existing model")
if __name__ == "__main__":
main()