Paper去污染数据集仍含52道HumanEval题：基准测试污染被低估

Paper

A research article published by the trifero.ai laboratory under Strife Technologies on data contamination in code benchmarks, which found that datasets labeled as decontaminated still contained 52 HumanEval problems.

Timeline (last 90 days)

Oct 1

The paper positions the research as a single-instance case study, with the core purpose of establishing a reproducible protocol that distinguishes conjecture, numerical verification, and proof.

Unverified50%

All Facts (1)

Unverified

The paper positions the research as a single-instance case study, with the core purpose of establishing a reproducible protocol that distinguishes conjecture, numerical verification, and proof.

50%

Source Articles

An LLM Call Is Really a Classifier: It Should Return a Label and a Probability, Not a SentenceOct 14 EgoOmniWorld: Unifying Egocentric Video Understanding via Ego-Action-World InteractionOct 14 Decontaminated Dataset Still Contains 52 HumanEval Problems: Benchmark Contamination Is UnderestimatedOct 13