Training Data & Post-TrainingYang et al.

Rethinking Benchmark and Contamination

Rethinking Benchmark and Contamination — Rephrased Samples

Why it matters

Rephrasing or translation alone defeats n-gram decontamination, and 8–18% of HumanEval overlaps with public corpora

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2311.04850

Added in Week 35, 2026 · See that week

Same week

OpenAI announces GPT-4 Turbo 128K and Assistants API RetrievalOfficialDevDayOfficial

Nearby on the same axis

NLP Evaluation in troublePaperDolmaPaperKTO — alignment from binary signals based on prospect theoryPaperDeepSeek introduces GRPO (DeepSeekMath)Paper