This repository contains a fine-tuned T5-small model designed to automate product metadata generation for e-commerce platforms. It simplifies the manual, time-consuming process of product categorization by predicting structured data like brand and hierarchical category levels (L0–L4) from basic product details.
- Base Model: T5-Small
- Dataset Link: Dataset
- Training: 1 epoch on ~440K examples
- Validation: ~60K examples
- Frameworks: PyTorch, Hugging Face Transformers, Datasets, Accelerate
For more training details refer to notebooks directory of refer which has two .ipynb notebooks for training and dataset creation (from existent datasource)
Manual product categorization is a resource-intensive task on e-commerce platforms.
Inputs:
- Product Title
- Manufacturer
- Store
Outputs:
- Brand
- L0 to L4 Category Levels (hierarchical classification)
The output is returned in a structured JSON format via a custom post-processing function.
An interactive Gradio app is included for testing and demo purposes. It integrates preprocessing and postprocessing pipelines.
🖼️ Gradio UI Preview
You can find and use the trained model on Hugging Face: 🔗 Hugging Face Model Link
- PyTorch
- transformers (Hugging Face)
- Datasets (Hugging Face)
- Accelerate (Hugging Face)
This project was a great learning journey—from fine-tuning and evaluation to deployment and UI creation. There's still room for improvement, and I welcome all feedback and contributions.
Feedback, issues, and pull requests are highly appreciated! Let’s make product metadata automation better together.