A template-and-combinatorial method produces HumanEval_T variants, and four LLMs score lower on them than on original HumanEval, suggesting likely data leakage.
Large language models for software engineering: Sur- vey and open problems,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.SE 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Addressing Data Leakage in HumanEval Using Combinatorial Test Design
A template-and-combinatorial method produces HumanEval_T variants, and four LLMs score lower on them than on original HumanEval, suggesting likely data leakage.