Research map / Large language models / Language, safety and interpretability
Detecting Data Contamination in LLMs via In-Context Learning
Michał Zawalski, Meriem Boubdir, Klaudia Bałazy, Besmira Nushi, Pablo Ribalta
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| detect contamination with CoDeC in-context scoring instead of vanilla loss, Min-K%, Zlib | all 13 models · seen vs unseen datasets | AUC: 75.7%, 78.5%, 89.6% -> 99.9% | Improved |
| finetune on a dataset, then measure CoDeC | Pythia, GPT-Neo, OLMo, Qwen3 · finetuning contamination | CoDeC score: consistently rose above 90% | Improved |
| finetune on MMLU, measure on cropped evaluation data | Pythia · MMLU (25% crop) | CoDeC score | No clear effect |
| finetune on MMLU, evaluate related QA datasets | Pythia · related QA benchmarks | CoDeC score: slightly elevated, far from high-contamination range | Mixed |
| finetune on MMLU, evaluate unrelated datasets | Pythia · Gutenberg books | CoDeC score | No clear effect |
| augment finetuning data (partial samples, augmented text) (ablation) | Pythia · MMLU | CoDeC score | No clear effect |
| add more in-context samples to the context (ablation) | CoDeC · Pythia 1.4B, 5 seen + 5 unseen datasets | seen/unseen separation | Improved |
| use 100 target examples (ablation) | CoDeC · Pythia 1.4B | score variance: 100 examples give stable estimates | No clear effect |
| use 1,000 target examples (ablation) | CoDeC · Pythia 1.4B | score variance: variance below 1% | Improved |
| CoDeC score evolution during OLMo 7B training (1k to 10k steps) | CoDeC · OLMo 7B | CoDeC score: contamination scores rise sharply between 1k and 10k steps (about 2% of training) | Improved |
| use higher finetuning learning rate on contaminated vs unseen data | finetuned models · contaminated and unseen datasets | target-token confidence | Mixed |
Closest papers by meaning
- Time Travel in LLMs: Tracing Data Contamination in Large Language Models (ICLR 2024 · similarity 0.93)
- Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination’s Impact on Machine Translation (ICML 2025 · similarity 0.91)
- How Much Can We Forget about Data Contamination? (ICML 2025 · similarity 0.91)
- ConStat: Performance-Based Contamination Detection in Large Language Models (NeurIPS 2024 · similarity 0.90)
- Task Contamination: Language Models May Not Be Few-Shot Anymore (AAAI 2024 · similarity 0.90)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
