CoSENT

Code for CoSENT: Consistent Sentence Embedding via Similarity Ranking

Overview

• main.py: main file for CoSENT.
• datasets: dictionary contains all datasets used in our paper.
• loss.py: CoSENT loss.
• arguments.py: configurations for training and evaluation.
• utils.py: tools for reading datasets.
• SentEval.py and senteval: codes for SentEval tasks, senteval is copied from SentEval: evaluation toolkit for sentence embeddings

Requirements

The implementation of CoSENT is tested under Python 3.9.13, with the following packages installed:

• pytorch==1.12.0
• tqdm==4.64.0
• transformers==4.23.1
• sentence-transformers==2.2.2
• datasets==2.6.1