siddeshrizwani/amazonml-price-prediction-transformer

This repository contains Team Helios's competition entry for the Amazon ML Challenge 2025, where they finished in the top 17 out of more than 3,000 competing teams. The task was to build a system that could make accurate predictions from messy, real-world product data: records with missing fields, inconsistent text, duplicate entries, and incomplete information. The team's approach moved through several steps. They first cleaned and normalized the raw data, then created new features to help the model learn better patterns. The core of the system uses transformer models, which are a type of neural network architecture well-suited to understanding text. They ran experiments with different settings, validated results carefully, and combined multiple models together at the end to improve the final score. The code is written in Python and uses PyTorch for building the models, along with Hugging Face Transformers for the pretrained language components. Data processing relied on standard libraries, and experiment results were tracked with Weights and Biases. Training required a GPU. The repository is organized into folders for data, notebooks, source code, saved models, and configuration files. Running the project involves installing the listed dependencies, then calling separate scripts for training and for generating predictions. The team published a detailed write-up of their approach on Medium for anyone who wants a deeper explanation of the competition strategy and key decisions.