Elamur Research⌕
☰
home / papers / 2023 / CLEAN-EVAL: Clean Evaluation on Contaminated Large Language Models

CLEAN-EVAL: Clean Evaluation on Contaminated Large Language Models

Paper · 2023 · Wenhong Zhu, Hongkun Hao, Zhiwei He et al.

Summary

Benchmark scores of large language models may be inflated when test items leak into pretraining data, and many developers do not disclose their training corpora. Labelling fresh test sets for each new model would avoid this but is slow and expensive. Clean-Eval instead rewrites existing benchmarks so that each item keeps its meaning but changes its surface form. An LLM paraphrases every contaminated sample, and the samples are also back-translated through several languages (for example English↔Chinese, German, and French). This produces a pool of candidates. A semantic detector discards candidates whose meaning has drifted. From the survivors, the method keeps the one with a moderate BLEURT score against the original, so it is reworded but still faithful.

The authors simulate contamination in two models, text-davinci-003 and Llama2, under in-context learning and fine-tuning. Across 20 benchmarks, the calibrated data pulls inflated scores back toward realistic levels. Human raters judged the rewritten items to be close to the originals in meaning. The approach offers a cheap way to re-check leaderboard claims without building new datasets.

Key points

  • Coverage: 20 benchmarks spanning arithmetic reasoning (GSM8K, MultiArith), examinations (MMLU, CEVAL), NLI and classification (SST2, MRPC, QNLI, QQP, SNLI, MNLI, CB, IMDB, AgNews), commonsense tasks (PIQA, COPA), and NLG (CNN/DailyMail, BBC XSUM).
  • Pipeline: LLM paraphrasing plus multi-language back-translation create a candidate set. A semantic-equivalence filter removes weak candidates, and BLEURT-based selection picks the final item.
  • Simulated contamination: In the in-context learning setup, the demonstration is identical to the test item. The clean comparison draws demonstrations from the train split. Clean-Eval tests on a reworded version of the demonstrated item.
  • Fine-tuning setup: Roughly 300 sampled instances per benchmark, with about 40 epochs of fine-tuning under a uniform configuration.
  • Stated limitation: Because the true training data of the evaluated models is unknown, the "no-contamination" baseline may itself overlap with pretraining data.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of EdinburghEdinburgh · United Kingdom359 works · 77,352 citations🎓Tsinghua UniversityBeijing · China333 works · 87,019 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Wenhong Zhu, Hongkun Hao, Zhiwei He, Yunze Song, Yumeng Zhang, Hanxu Hu, Yiran Wei, Rui Wang and Hongyuan Lu. 2023. CLEAN-EVAL: Clean Evaluation on Contaminated Large Language Models. arXiv:2311.09154.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2311.09154 · arXiv 2311.09154

@misc{zhu2023clean,
  title = {{CLEAN-EVAL: Clean Evaluation on Contaminated Large Language Models}},
  author = {Wenhong Zhu, Hongkun Hao, Zhiwei He, Yunze Song, Yumeng Zhang, Hanxu Hu, Yiran Wei, Rui Wang and Hongyuan Lu},
  year = {2023},
  doi = {10.48550/arxiv.2311.09154},
  eprint = {2311.09154},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2311.09154}
}

References

15 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Zejun Li, Ye Wang, Mengfei Du, Qingwen Liu, Binhao Wu, Jiwen Z… find
  2. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics.
  3. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon…
  4. Inbal Magar and Roy Schwartz. 2022. Data contamination: From memorization to exploitation. find
  5. Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797–1807, Brussels, Belgium. As…
  6. OpenAI. 2023. Gpt-4 technical report.
  7. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073– 1083, Vancouver, Canada. Association for Compu… find
  8. Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881–7892, Online. Association for Computational Linguistics.
  9. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86–96, Berlin, Germany. Association for Computational L… find
  10. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller,… — Llama 2: Open Foundation and Fine-Tuned Chat Models
  11. Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2019a. Superglue: A stickier benchmark for general-purpose language understanding systems. arXiv preprint arXiv:1905.00537.
  12. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019b. Glue: A multi-task benchmark and analysis platform for natural language understanding.
  13. Andy Way. 2013. Emerging use-cases for machine translation. In Proceedings of Translating and the Computer 35. find
  14. Tianwen Wei, Liang Zhao, Lichang Zhang, Bo Zhu, Lijie Wang, Haihua Yang, Biye Li, Cheng Cheng, Weiwei Lü, Rui Hu, Chenxia Li, Liu Yang, Xilin Luo, Xuejie Wu, Lunan Liu, Wenjun Cheng, Peng Cheng, Jianhao Zhang, Xiaoyu Zhang, Lei Lin, Xiaokun Wang, Yutuan Ma, Chuanhai Dong, Yanqi Sun, Yifu Chen, Yongy… — Skywork: A More Open Bilingual Foundation Model
  15. Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In NIPS.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2023-clean-eval-clean-evaluation-contaminated/