env-audit-contamination

Check 6 — contamination

Question: does this env's dataset overlap the datasets the user cares about, so that "improvement" measured on them would be partly memorization?

Runs only against user-provided datasets. The user names the HuggingFace datasets to check (ids like openai/gsm8k, or hf.co links — normalize a link to its org/name id). Do not pick benchmarks yourself. If the user provided none, output immediately:

{"name": "contamination", "status": "N/A", "score": null,
 "justification": "no contamination datasets provided"}

(N/A is excluded from the rating, so the check carries no weight when skipped.)

Steps

1. Load each provided dataset with the datasets library (it ships with verifiers), e.g. python -c "from datasets import load_dataset; ds = load_dataset('openai/gsm8k', 'main', split='test')". Use the split the user named; default to test, falling back to train. If a dataset fails to load because it is gated / needs credentials (error mentions gated, 401/403, authentication, token), don't skip it silently — ask the user for an HF token (and access approval) first, set HF_TOKEN, and retry; only proceed without it if they decline. If a dataset can't be loaded for other reasons (bad id, offline), say so in the justification and judge from the ones that did load; if none load, output N/A with that reason.

2. Get a good slice of the env's dataset. Re-run rlenv-audit inspect <env> -n 100 (or more) so the comparison isn't limited to the orchestrator's 20-sample inspect.

3. Compare instances. For each env instance vs. each provided dataset:

   • normalize the question text (lowercase, collapse whitespace, strip punctuation) and look for exact or near-exact matches (high n-gram containment), not just shared boilerplate;
   • catch paraphrases the n-grams miss: for near-misses, compare the numbers, named entities, and the answer — a reworded question with the same figures and the same answer is the same instance. Judge a sample of suspicious pairs yourself rather than trusting text overlap alone;
   • distinguish same-template-different-instance (fine) from same instance (contamination);
   • check both splits of each provided dataset where they exist — env tasks lifted from the benchmark's train split still poison eval-set conclusions when the splits share templates;
   • report concrete matching pairs (env row ↔ dataset row) with counts;
   • note the train/eval distinction: an explicit eval env that is one of the provided datasets is expected — say so; a training dataset overlapping the user's eval set is the real problem.

Output

Score 0–10 where 10 = clean, lower = more overlap. Anchor the score to the measured rate over the sampled rows: no same-instance matches → 9–10; isolated matches (≲5% of the sample) → 6–8 (WARN); systematic overlap (tens of percent) → ≤5; the env is the user's eval set used for training → FAIL territory. State the sample size — 0 matches in 100 rows is evidence, not proof.

Any number of datasets may be provided — audit every one independently and score on the worst offender: four clean datasets must not dilute one contaminated one. The justification names each dataset with its match count (e.g. openai/gsm8k: 3 near-matches; hendrycks/MATH: clean; ...).

{"name": "contamination", "status": "PASS|WARN|FAIL", "score": <0-10>,
 "justification": "<one line: datasets checked, matches found (counts) or clean>"}