# One Embedder, Any Task: Instruction-Finetuned Text Embeddings

**Authors:** Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, Tao Yu  
**arXiv:** [2212.09741](https://arxiv.org/abs/2212.09741)  
**Date:** 2022-12-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.09741v1/

## Key points

- INSTRUCTOR is a single embedder that generates text embeddings based on task instructions
- Trained using multitask mixture with contrastive loss on annotated instructions for 330 diverse tasks
- Achieves state-of-the-art performance across diverse datasets with an average improvement of 3.4%
- Robust to changes in instructions and effective in addressing the challenge of training a single model on varied datasets
- Utilizes text embeddings as features for classifiers trained on training data while keeping sentence embeddings frozen for classification tasks
- Measures similarity between sentence pairs through cosine similarity of embeddings for semantic textual similarity (STS) tasks
- Discusses automatic summarization evaluation methods in the context of evaluating performance on downstream tasks

## Summary

The paper "One Embedder, Any Task: Instruction-Finetuned Text Embeddings" introduces the innovative method of INSTRUCTOR for generating text embeddings based on task instructions. Unlike previous specialized encoders, INSTRUCTOR is a single embedder that can produce customized text embeddings for various downstream tasks and domains without additional training. The researchers annotated instructions for 330 diverse tasks and trained INSTRUCTOR using a multitask mixture with contrastive loss. The evaluation of INSTRUCTOR on 70 embedding tasks covers classification, information retrieval, semantic textual similarity, and text generation evaluations. Despite having fewer parameters than the previous best model, INSTRUCTOR achieves state-of-the-art performance with an average improvement of 3.4% across diverse datasets. The analysis shows that INSTRUCTOR is robust to changes in instructions and that instruction finetuning effectively addresses the challenge of training a single model on varied datasets. Additionally, the paper discusses how INSTRUCTOR performs on classification tasks by using text embeddings as features for classifiers trained on training data while keeping sentence embeddings frozen. It also explores semantic textual similarity (STS) tasks where the model measures similarity between sentence pairs through cosine similarity of their embeddings. Furthermore, automatic summarization evaluation methods are discussed in the context of evaluating performance on downstream tasks. Overall, the study showcases the versatility and effectiveness of INSTRUCTOR in generating high-quality text embeddings tailored to different tasks and domains without requiring extensive retraining or specialization.