efficient splade V I B T large doc

Introduction

The Efficient SPLADE model is designed for passage retrieval, employing separate models for query and document inference. It focuses on enhancing efficiency and performance in information retrieval tasks, particularly using the MS MARCO dataset.

Architecture

Efficient SPLADE uses a dual-model architecture, with distinct models for query and document processing. This separation allows for optimized handling of each type of input. The document model is available at Efficient SPLADE VI-BT Large Doc, and the corresponding query model can be found here.

Training

The model was trained using various optimization techniques, including L1 regularization for queries, separation of document/query encoders, FLOPS-regularized middle training, and faster query encoders. These methods aim to improve efficiency and reduce latency, achieving competitive performance metrics such as MRR@10 and R@1000 on the MS MARCO development dataset.

Guide: Running Locally

To run the Efficient SPLADE model locally, follow these steps:

1. Clone the Repository: Clone the SPLADE code repository from GitHub.
2. Install Dependencies: Ensure all necessary dependencies are installed, typically using a package manager like pip.
3. Download Models: Obtain both the document and query models from Hugging Face.
4. Run Inference: Use a script from the repository to load the models and perform inference on your data.

For optimal performance, consider using cloud GPUs from providers like AWS, Google Cloud, or Azure to handle computational demands.

License

Efficient SPLADE is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (cc-by-nc-sa-4.0). This permits adaptation and sharing under similar terms, but not for commercial use.