ProbDiff

Code of paper "Language Models can Evaluate Themselves via Probability Discrepancy"

Overview

We introduce ProbDiff, a self-evaluation technique applicable to any LLM across tasks. Given a query $q$, ProbDiff first prompts a candidate LLM to generate a response $x$, then asks the LLM to revise $x$ based on $q$, producing a refined response $\hat{x}$. Finally, ProbDiff quantifies the probability discrepancy ${\rm log};p({\hat x}|q)$ - ${\rm log}; p(x|q)$ as the evaluation metric. When comparing two candidate LLMs on $q$, a larger probability discrepancy indicates a lower proficiency in handling the instruction.

Environment

Framework Versions:

python=3.10
torch=2.1.2
transformers=4.37.2
vllm=0.3.0

Data

Xiaohongshu Blog writting dataset is stored in generation_task/data/redbook

Scripts

We provide xxx.sh to reproduce the results of ProbDiff in each folder.

Citation

If you finding our work interesting or helpful to you, please cite this repo.

@inproceedings{xia-etal-2024-language,
    title = "Language Models can Evaluate Themselves via Probability Discrepancy",
    author = "Xia, Tingyu  and Yu, Bowen  and Wu, Yuan  and Chang, Yi  and Zhou, Chang",
    booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
    month = aug,
    year = "2024",
    publisher = "Association for Computational Linguistics",
}