Amazon Price Predictor 🚀

Ever wondered how much that shiny new gadget on Amazon really costs? What if an AI could guess the price just from the product description? Enter Amazon Price Predictor – a fine-tuned Llama model that turns product blurbs into dollar signs! Built on Meta's Llama 3.1-8B, this project fine-tunes the beast for spot-on price predictions, all while keeping things lightweight and efficient. No more overpaying – let AI do the haggling! 💸

Why is this cool? In a world of endless e-commerce listings, accurate pricing AI can power recommendation engines, detect deals, or even automate inventory valuation. And guess what? We did it on a budget using Kaggle's free GPU quota – proving you don't need a supercomputer to build something awesome.

Key Features ✨

• Smart Predictions: Predicts prices from product descriptions with impressive accuracy.
• Lite Dataset: Focused on Home Appliances from Amazon Reviews 2023 – compact yet effective for quick training.
• Efficient Fine-Tuning: Uses QLoRA to train on limited hardware without sacrificing performance.
• Visual Insights: Training monitored on Weights & Biases (WandB), results visualized with error metrics and charts.
• Error Analysis: Color-coded predictions (green for spot-on, orange for close, red for misses) make debugging fun.

Dataset: The Foundation 📊

We curated a "lite" dataset from the massive Amazon Reviews 2023 corpus, zeroing in on Home Appliances to keep things snappy.

• Why Appliances? Everyday items like fridges and toasters provide diverse descriptions and prices, perfect for a focused model without bloating compute needs.
• Processing Magic: Loaded via Hugging Face, filtered for valid prices/descriptions, tokenized with Llama's tokenizer, and split into 10k train + 2k test samples.
• Stats at a Glance: Average token length ~100, prices ranging from $0 to $1000+.

Here's a peek at the dataset distribution:

Uploaded to Hugging Face as ishant24/lite-data for easy access. Custom scripts in items.py and loaders.py handle the heavy lifting – check data_prep.ipynb for the full prep workflow.

Training: From Base to Beast 🛠️

We fine-tuned Llama 3.1-8B using QLoRA (Quantized Low-Rank Adaptation) – a memory-efficient technique that adapts the model without full retraining.

• Setup: Ran on Kaggle's NVIDIA Tesla T4 GPU (30 free hours/week – budget-friendly!).
• Data: 10k training samples (to keep costs low; full 25k works too but takes longer).
• Hyperparams: 1 epoch (avoids overfitting; loss drops steadily), batch size 6, learning rate 1e-4, cosine scheduler.
• Monitoring: WandB for real-time loss graphs and gradients – because who doesn't love watching training curves dip?

Training graph from WandB:

The process is detailed in training.ipynb. We used libraries like Transformers, PEFT, and TRL for seamless SFT (Supervised Fine-Tuning).

Testing: Proof in the Predictions 📈

Evaluation happens in Test.ipynb, where we pit the model against unseen test data.

• Metrics: Mean Absolute Error (MAE) for dollar differences, RMSLE for relative accuracy.
• Visualizer: Custom Tester class runs predictions, colors errors (green: < $40 or 20%, orange: < $80 or 40%, red: oops!), and plots truths vs. guesses.
• Results: On 250 test samples, MAE $18.87, with ~90.4% "hits" (green). Far Far better from base model!

Results graph:

Why so good? The model captures nuances like brand, features, and quality from descriptions – e.g., "premium stainless steel fridge" gets a higher guess than "basic mini cooler." Errors often stem from outliers (super-cheap or luxury items), but overall, it's a solid e-commerce sidekick.

Why This Model Rocks 🎉

• Efficient & Accessible: Trains in hours on free resources – democratizing AI for hobbyists.
• Accurate Enough for Real Use: Low MAE means reliable estimates; RMSLE handles price scales well.
• Scalable: Lite version focuses on appliances, but swap datasets for electronics, toys, etc.
• Fun Insights: Spot overpriced listings or predict sales – endless applications!

In benchmarks (on our test set), it outperforms base Llama (which hallucinates prices) and simple baselines like average price guessing.

Setup & Installation 🛠️

1. Clone the repo: git clone https://github.com/yourusername/amazon-price-predictor.git
2. Install deps: pip install -r requirements.txt (or from notebooks).
3. Set up .env with HF_TOKEN for Hugging Face access.
4. Run data_prep.ipynb to prep data (or use pre-uploaded dataset).

Usage 🚀

• Prep Data: Run data_prep.ipynb to generate train/test sets.
• Train: Fire up training.ipynb on Kaggle/GPU.
• Test: Load model in Test.ipynb and evaluate with Tester.test(predict_fn, test_data).