Skip to content
/ SAR Public

[ACL 2024] Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models

License

Notifications You must be signed in to change notification settings

jinhaoduan/SAR

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models [ACL 2024]

arXiv

Authors: Jinhao Duan, Hao Cheng, Shiqi Wang, Chenan Wang, Alex Zavalny, Renjing Xu, Bhavya Kailkhura, Kaidi Xu

The proposed Shifting-Attention-to-Relevance (SAR) is implemented in this codebase.

Updates

[8/2024] 🎉🎉 Glad to know that SAR is ranked 1st among 28 LLM uncertainty quantification methods in LM-Polygraph. Please also check their implementation and paper.

Environments

Please config environment by

pip install -r requirements.txt

Data Preparing

cd src
sh parse_datasets.sh

It will automatically parse CoQA, Trivia QA, and SciQ datasets.

Uncertainty Estimation for Open-source LLMs

for the CoQA dataset

sh scripts/coqa/ue_pipeline_opt-2.7b.sh

sh scripts/coqa/ue_pipeline_opt-6.7b.sh

sh scripts/coqa/ue_pipeline_opt-13b.sh

sh scripts/coqa/ue_pipeline_opt-30b.sh

sh scripts/coqa/ue_pipeline_llama-7b.sh

sh scripts/coqa/ue_pipeline_llama-13b.sh

for the SciQ dataset:

sh scripts/sciq/ue_pipeline_opt-2.7b.sh

sh scripts/sciq/ue_pipeline_opt-6.7b.sh

sh scripts/sciq/ue_pipeline_opt-13b.sh

sh scripts/sciq/ue_pipeline_opt-30b.sh

sh scripts/sciq/ue_pipeline_llama-7b.sh

sh scripts/sciq/ue_pipeline_llama-13b.sh

for the Trivia QA dataset:

sh scripts/trivia_qa/ue_pipeline_llama-7b.sh

sh scripts/trivia_qa/ue_pipeline_llama-13b.sh

Reference

Please cite our paper if you feel this is helpful:

@inproceedings{duan2024shifting,
  title={Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models},
  author={Duan, Jinhao and Cheng, Hao and Wang, Shiqi and Zavalny, Alex and Wang, Chenan and Xu, Renjing and Kailkhura, Bhavya and Xu, Kaidi},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={5050--5063},
  year={2024}
}

Acknowledgement

This codebase is build upon Semantic Entropy (SE). Thanks for their excellent contribution!

About

[ACL 2024] Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published