Collu-Bench
收藏资源简介:
Collu-Bench是由普渡大学创建的一个用于预测语言模型在代码中产生幻觉的基准数据集。该数据集包含13,234个代码幻觉实例,收集自五个数据集和11个不同的大型语言模型。数据集内容丰富,包括每一步的日志概率、Token类型和执行反馈等详细特征,旨在帮助研究人员深入分析代码幻觉。创建过程中,通过自动化管道处理程序等价性和标识符可行性,确保幻觉Token位置的准确性。Collu-Bench主要应用于代码生成和自动化程序修复任务,旨在解决大型语言模型在代码生成过程中产生的幻觉问题,提高代码的准确性和可靠性。
Collu-Bench is a benchmark dataset developed by Purdue University for predicting hallucinations in code generated by language models. This dataset comprises 13,234 code hallucination instances collected from five datasets and 11 distinct large language models. It includes rich detailed features such as step-by-step log probabilities, Token types, and execution feedback, designed to assist researchers in conducting in-depth analyses of code hallucinations. During its development, an automated pipeline was employed to handle program equivalence and identifier feasibility, ensuring the accuracy of the positions of hallucinatory Tokens. Collu-Bench is primarily applied to code generation and automated program repair tasks, aiming to address the hallucination issues arising during code generation by large language models and improve the accuracy and reliability of generated code.

- 1Collu-Bench: A Benchmark for Predicting Language Model Hallucinations in Code普渡大学 · 2024年



