Pricer

Pricer is a project that demonstrates fine-tuning a large language model (Meta-Llama 3.1-8B) using QLoRA techniques to predict item prices based on textual descriptions. The project leverages custom data processing, efficient LoRA-based fine-tuning, and evaluation scripts to produce a model that estimates prices from item details.

AIM:

Fine-tune LLaMA 3.1 using QLoRA for accurate price prediction from textual product descriptions, outperforming frontier models like GPT-4o and GPT-4o Mini

Results

Comparison Table

Model	Error	RMSLE	Hits
GPT 4o	74.53	0.0317	56.0%
GPT 4o Mini	87.98	0.0594	50.0%
LLaMA 3.1	395.75	1.49	28.0%
Fine-Tuned LLaMA 3.1	64.06	0.49	58.0%

Below are the model performance results with visualizations.

GPT-4o	GPT-4o Mini
LLAMA-3.1 Without Fine-Tuning	LLAMA-3.1 Fine-Tuned

Features

• QLoRA Fine-Tuning: Uses LoRA (Low-Rank Adaptation) to fine-tune the Meta-Llama 3.1-8B model efficiently.
• Custom Data Processing: Implements data cleaning and prompt generation for price prediction using custom modules.
• Parallel Data Loading: Uses parallel processing to load and preprocess data from the Amazon Reviews dataset.
• Model Evaluation: Provides testing scripts to assess model performance and visualize prediction errors.
• HuggingFace Hub Integration: Pushes the fine-tuned model to the HuggingFace Hub for easy sharing and deployment.
• Weights & Biases Tracking (Optional): Integrates with Weights & Biases for experiment tracking and monitoring.

Project Structure

• qlora_llama_3_1_finetune.py
  Main script for fine-tuning the Llama 3.1 model using QLoRA. It configures the training parameters, LoRA settings, and dataset details, and pushes the trained model to the HuggingFace Hub.
  :contentReference[oaicite:0]{index=0}

• items.py
  Contains the Item class, which is responsible for cleaning raw item data, generating prompts, and preparing text for training.
  :contentReference[oaicite:1]{index=1}

• loaders.py
  Implements the ItemLoader class that loads and processes the dataset in parallel from the "McAuley-Lab/Amazon-Reviews-2023" dataset.
  :contentReference[oaicite:2]{index=2}

• tester.py
  Provides tools for evaluating model performance. It includes functionality to compute errors, plot predictions versus ground truth, and generate reports based on the model’s estimates.
  :contentReference[oaicite:3]{index=3}

• Additional Notebooks:
  Other notebooks (e.g., for dataset generation, baseline models, testing unfine-tuned models) are included to support different stages of the development and evaluation pipeline.