Benchmark Contamination › Real-World Contamination Discoveries

Several high-profile contamination discoveries have shaped the field's practices and awareness. Reviewing them builds intuition for how contamination manifests in practice.

The GPT-3 paper acknowledged benchmark contamination explicitly when reporting results, noting that WebText (the training corpus) likely contained overlap with several benchmarks. The paper included a contamination analysis showing performance drops when contaminated examples were removed, and reported results both with and without the contaminated subset. This transparency set a standard that subsequent papers have followed to varying degrees.

The Llama 2 technical report found that some proportion of common benchmark examples appeared in its training data sourced from public internet crawls. The authors performed n-gram analysis at $n=13$ and reported contaminated benchmark subsets alongside clean-only performance figures. They found that contaminated subsets did show higher accuracy in most cases, confirming that contamination artificially inflates scores and that removing contaminated examples is necessary for fair reporting.

The ChatGPT and GPT-4 evaluations faced contamination scrutiny because OpenAI's training data is not fully disclosed. Researchers outside OpenAI used differential perplexity analysis and other black-box methods to estimate contamination levels, finding evidence of exposure to several benchmarks. The lack of training data transparency makes definitive contamination analysis impossible, which is itself a form of evaluation opacity that the field has struggled to address.

BIG-bench, a collaborative benchmark across many research groups, specifically designed canary strings and per-example difficulty metadata to support contamination analysis. The benchmark's design philosophy acknowledged that static benchmarks inevitably get contaminated and built infrastructure for detecting and quantifying this contamination over time.