efficient splade V large doc

Introduction

Efficient SPLADE is a model designed for passage retrieval using a distinctive architecture that employs separate models for query and document inference. This document specifically pertains to the document model. The Efficient SPLADE model is optimized for efficiency and performance in information retrieval tasks.

Architecture

The Efficient SPLADE architecture involves two distinct models: one for queries and another for documents. This separation allows for optimized processing and retrieval efficiency. The model utilizes techniques such as L1 regularization, separation of encoder roles, and FLOPS-regularized training to enhance performance while maintaining low latency, comparable to traditional retrieval systems like BM25.

Training

The model was trained using the MS MARCO dataset, optimizing for metrics such as MRR@10 and R@1000. The architecture allows for adjustments in efficiency through regularization factors, which aim to balance performance with computational demands.

Guide: Running Locally

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

1. Set Up Environment: Install PyTorch and Hugging Face's Transformers library.
2. Download Model: Obtain the document and query models from their respective Hugging Face model pages.
3. Load Dataset: Use the MS MARCO dataset or another relevant dataset for input data.
4. Run Inference: Execute the models for document and query inference as intended.

For enhanced performance, consider using cloud GPU services such as AWS EC2, Google Cloud Platform, or Azure for training and inference.

License

The Efficient SPLADE model is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (cc-by-nc-sa-4.0). This allows for sharing and adaptation under specified conditions, primarily for non-commercial use.