Skip to content

🧠 Machine Learning Models

This section details the models used for price prediction in Snapr.

Fine-Tuned LLaMA Model

We fine-tuned a LLaMA 3.1 8B Quantized model using QLoRA on a curated subset (~400K items) from a larger dataset.

🔗 See Data Pipeline for dataset source and preprocessing steps.

After fine-tuning and evaluation, the best-performing checkpoint was published to Hugging Face Hub.

📓 Related Notebooks:

At runtime, Modal pulls the model from Hugging Face on first use, caches it, and uses it for zero-shot product price prediction.

XGBoost Model

We reused the same curated dataset to generate embeddings using the E5 model, and stored them in a ChromaDB, which was uploaded to AWS S3 for remote access.

Using these embeddings, we trained an XGBoost regression model with tuned hyperparameters to predict product prices from vectorized descriptions in the training dataset. The trained model (xgboost_model.pkl) was pushed to Hugging Face Hub.

The implementation details are documented in the following notebook

At runtime, Modal pulls:

  • the XGBoost model from Hugging Face,
  • the E5 embedding model (from sentence-transformers on Hugging Face),
  • and the ChromaDB from AWS S3.

All components are cached in Modal’s persistent volume after first use for fast, repeatable access.

RAG Pipeline

We also use the ChromaDB with E5 embeddings to power a RAG pipeline for price prediction.

  • Retrieval: Given a product description, we embed it using the E5 model and retrieve the top 5 most similar items from ChromaDB. Each retrieved item includes its description and actual price.

  • Augmented: The retrieved similar items and their prices are combined with the original product description to form the input context, which augments the prompt sent to the language model for price prediction.

  • Generation: The full prompt — containing the product description and similar item data — is sent to GPT via the OpenAI API. The model is instructed to output only the estimated price.

Ensemble Model

After obtaining individual price predictions from the three independent models on the curated dataset, we trained a linear regression model to combine their outputs.

The model uses the raw predictions along with simple engineered features (e.g., max, mean) to generate a more stable final estimate.

Once trained and evaluated, the ensemble model (ensemble_model.pkl) was pushed to Hugging Face Hub.

📓 Related Notebook

At runtime, Modal pulls the model from Hugging Face (on first use) and caches it in a persistent volume for efficient reuse during future predictions.

Frontier Model (OpenAI)

We use OpenAI GPTto:

  • Select the top 5 deals from raw inputs, ensuring clear descriptions and explicit prices, and
  • Estimate product prices after retrieving similar items from ChromaDB for contextual grounding via RAG.