Simulating Training Data Leakage in Multiple-Choice Benchmarks for LLM Evaluation
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910975199805440 |
|---|---|
| author | Hidayat, Naila Shafirni Kautsar, Muhammad Dehan Al Wicaksono, Alfan Farizki Koto, Fajri |
| author_facet | Hidayat, Naila Shafirni Kautsar, Muhammad Dehan Al Wicaksono, Alfan Farizki Koto, Fajri |
| contents | The performance of large language models (LLMs) continues to improve, as reflected in rising scores on standard benchmarks. However, the lack of transparency around training data raises concerns about potential overlap with evaluation sets and the fairness of reported results. Although prior work has proposed methods for detecting data leakage, these approaches primarily focus on identifying outliers and have not been evaluated under controlled simulated leakage conditions. In this work, we compare existing leakage detection techniques, namely permutation and n-gram-based methods, under a continual pretraining setup that simulates real-world leakage scenarios, and additionally explore a lightweight method we call semi-half question. Although semi-half offers a low-cost alternative, our analysis shows that the n-gram method consistently achieves the highest F1-score. We also refine these techniques to support instance-level detection and reduce computational overhead. Leveraging the best-performing method, we create cleaned versions of MMLU and HellaSwag, and re-evaluate several LLMs. Our findings present a practical path toward more reliable and transparent evaluations, and we recommend contamination checks as a standard step before releasing benchmark results. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_24263 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Simulating Training Data Leakage in Multiple-Choice Benchmarks for LLM Evaluation Hidayat, Naila Shafirni Kautsar, Muhammad Dehan Al Wicaksono, Alfan Farizki Koto, Fajri Computation and Language The performance of large language models (LLMs) continues to improve, as reflected in rising scores on standard benchmarks. However, the lack of transparency around training data raises concerns about potential overlap with evaluation sets and the fairness of reported results. Although prior work has proposed methods for detecting data leakage, these approaches primarily focus on identifying outliers and have not been evaluated under controlled simulated leakage conditions. In this work, we compare existing leakage detection techniques, namely permutation and n-gram-based methods, under a continual pretraining setup that simulates real-world leakage scenarios, and additionally explore a lightweight method we call semi-half question. Although semi-half offers a low-cost alternative, our analysis shows that the n-gram method consistently achieves the highest F1-score. We also refine these techniques to support instance-level detection and reduce computational overhead. Leveraging the best-performing method, we create cleaned versions of MMLU and HellaSwag, and re-evaluate several LLMs. Our findings present a practical path toward more reliable and transparent evaluations, and we recommend contamination checks as a standard step before releasing benchmark results. |
| title | Simulating Training Data Leakage in Multiple-Choice Benchmarks for LLM Evaluation |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2505.24263 |