Public Benchmarks & Contamination
AI Engineer path · lesson 91 of 124Read a leaderboard properly. Compare two suites that share two benchmarks out of ten, spot which ones correlate at 0.90 and are counting the same capability twice, and watch a headline score inflate while the clean-samples score never moves.
Key insight
Ask what a leaderboard left out before you ask what it ranked. The benchmarks that appear are the ones that are cheap, automatic and public — which systematically under-represents retrieval, long-context work, and anything that needs a human or a sandbox to score.
Detect a contaminated benchmark
Order a contamination check.
Tap a piece, then tap where it goes — or drag it. Tap the tray to take one back.
- 1Drop a step here
- 2Drop a step here
- 3Drop a step here
- 4Drop a step here
- 5Drop a step here
Test your knowledge
Roughly how many benchmarks does Google's BIG-bench contain?
The short version
A school that publishes its league-table position. The exams are public, they are marked by the school, several of them test the same subject under different names, and last year's papers are freely available to revise from. None of that makes the score meaningless. It does mean you should read it as evidence rather than as the answer.
In more depth
Public Benchmarks & Contamination
There are more benchmarks than anyone can run: BIG-bench alone has 214, EleutherAI's harness supports 400+, OpenAI's evals about 500. Every leaderboard is a compute-constrained, opinionated selection.
The leaderboards disagree
Hugging Face Open LLM Leaderboard, late 2023 — six benchmarks: ARC-C, MMLU (57 subjects), HellaSwag, TruthfulQA, WinoGrande, GSM-8K.
Stanford HELM used ten benchmarks, only two of which (MMLU and GSM-8K) overlapped. The other eight: MATH, LegalBench, MedQA, WMT 2014, NarrativeQA, OpenBookQA, and Natural Questions under two settings.
June 2024 revamp (under a year later): GSM-8K → MATH lvl 5, MMLU → MMLU-PRO, plus GPQA, MuSR, BBH, IFEval.
Leaderboards are compute-constrained: HELM Lite left out MS MARCO; Hugging Face opted out of HumanEval. Stanford spent $80,000–$100,000 to evaluate 30 models on the full HELM suite.
Redundancy — Pearson correlations (Jan 2024)
| Pair | r |
|---|---|
| MMLU ↔ WinoGrande | 0.9011 |
| ARC-C ↔ WinoGrande | 0.8856 |
| ARC-C ↔ MMLU | 0.8672 |
| MMLU ↔ GSM-8K | 0.7936 |
| HellaSwag ↔ GSM-8K | 0.3547 |
| TruthfulQA ↔ HellaSwag | 0.4228 |
WinoGrande, MMLU and ARC-C are strongly correlated — all test reasoning. TruthfulQA is only moderately correlated: improving reasoning and maths doesn't always improve truthfulness.
Aggregation
The same four scores can rank two models in opposite orders under the two methods.
Contamination
Also called data leakage, training on the test set, or cheating. Schaeffer's satirical "Pretraining on the Test Set Is All You Need" (2023): a one-million-parameter model trained only on benchmark data scored near-perfectly and outperformed much larger models.
Three kinds: unintentional (web scraping), indirect (same source material), and intentional-for-good-reasons (benchmark data is high-quality data).
Two detections: n-gram overlap — a shared sequence of 13 tokens marks a sample dirty; accurate, expensive, needs training-data access. Perplexity — unusually low perplexity suggests the model saw it; less accurate, far cheaper, works from outside.
OpenAI's GPT-3 analysis found 13 benchmarks with at least 40% of their data in the training set.
"A benchmark stops being useful as soon as it becomes public."
Interview Questions
References & Further Reading
Love this visualization?
Sign up to track your progress, earn achievements, use the AI assistant, and access your personal learning dashboard.
100% free • No credit card required