Uncertainty-driven Embedding Convolution

This repository contains PyTorch implemenations of "Uncertainty-driven Embedding Convolution" .

Introduction

Uncertainty-driven Embedding Convolution (UEC) is a framework for combining embeddings in principled, uncertainty-aware manner. UEC consists of three key components.

1. post-hoc conversion of deterministic embeddings into probabilistic ones via Laplace approximation
2. Gaussian convolution with uncertainty-aware weights
3. uncertainty-aware similarity scoring

🔧 Installation

To install the required dependencies, please follow the steps in order:

# 1. Install the base dependencies
pip install -r requirements.txt

# 2. Manually install flash-attn (version must match)
pip install flash-attn==2.7.4.post1

▶️ Running Scripts

Below are example commands to execute the different Python scripts in this project.

1. Multilingual Evaluation

For details on how to run the multilingual evaluation, please refer to the Multilingual Evaluation README.

2. MTEB Experiments

For details on how to run the MTEB experiments, please refer to the MTEB Experiments README.