Trove documentation

Trove is a flexible toolkit for training and evaluating dense retrievers. It aims to keep the codebase simple and hackable, while offering a clean, unified interface for quickly experimenting with new ideas.

Repo: BatsResearch/trove


Trove Logo

Install Trove from PyPI:

pip install ir-trove

To get the latest changes, install from source:

pip install git+https://github.com/BatsResearch/trove

To get started with Trove, explore the following resources:

  • README: General overview of Trove

  • Training: Step-by-step guide for training dense retrievers with Trove

  • Inference: Step-by-step guide for evaluation, hard negative mining, and encoding

  • Examples: Collection of self-contained scripts for training and inference

  • Data: Detailed guide for loading, preprocessing, and managing datasets

  • Modeling: Insight into Trove’s modeling architecture and how to extend it with custom models, loss functions, and more