Deciphering the Impact of Pretraining Data on Large Language Models through Machine Unlearning
Open accessAbstract
Through pretraining on a corpus with various sources, Large Language Models (LLMs) have gained impressive performance.However, the impact of each component of the pretraining corpus remains opaque.As a result, the organization of the pretraining corpus is still empirical and may deviate from the optimal.To address this issue, we systematically analyze the impact of 48 datasets from 5 major categories of pretraining data of LLMs and measure their impacts on LLMs using benchmarks about nine major categories of model capabilities.Our analyses provide empirical results about the contribution of multiple corpora on the performances of LLMs, along with their joint impact patterns, including complementary, orthogonal, and correlational relationships.We also identify a set of "high-impact data" such as Books that is significantly related to a set of model capabilities.These findings provide insights into the organization of data to support more efficient pretraining of LLMs.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
LLM training data includes vast quantities of Internet-scraped text likely to include medical misinformation, health advice written by laypeople, racist and sexist content, and outdated guidelines.
“LLM training data includes not only peer-reviewed scientific literature but also vast quantities of Internet-scraped text likely to include medical misinformation, health advice written by laypeople, racist and sexist content, and outdated guidelines4.”
From Large language models provide unsafe answers to patient-posed medical questions (2026) · cites this paper as [4]
Section 3.1.1 (page unavailable) identifies ArXiv alongside C4, Wikipedia, and StackExchange as components of an LLM pretraining corpus, supporting the broad point that training data comes from varied sources beyond scientific literature. It does not establish the quantity of Internet-scraped text or address medical misinformation, layperson health advice, racist or sexist content, or outdated guidelines. (LLM, conf=0.87)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 21
- Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. Drop: A reading comprehension benchmark requir- ing discrete reasoning over paragraphs. In North American Chapter of the Association for Computa- tional Linguistics.
- Paul Ernest. 2023. Abduction and creativity in math- ematics. Handbook of Abductive Cognition, pages 585-611.
- Amirata Ghorbani and James Zou. 2019. Data shapley: Equitable valuation of data for machine learning. In International conference on machine learning, pages 2242-2251. PMLR.
- Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. 2021. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 34:16319-16330.
- Satoshi Hara, Atsushi Nitanda, and Takanori Maehara. 2019. Data cleansing for models trained with sgd. Advances in Neural Information Processing Systems, 32.
- Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2021a. Aligning ai with shared human values. Pro- ceedings of the International Conference on Learning Representations (ICLR).
- Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Stein- hardt. 2021b. Measuring massive multitask language understanding. Proceedings of the International Con- ference on Learning Representations (ICLR).
- Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021c. Measuring mathematical problem solving with the math dataset. NeurIPS.
- Ruoxi Jia, Fan Wu, Xuehui Sun, Jiacen Xu, David Dao, Bhavya Kailkhura, Ce Zhang, Bo Li, and Dawn Song. 2021. Scalability vs. utility: Do we have to sacrifice one for the other in data importance quantification? In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 8239- 8247.
- Nikhil Kandpal, Eric Wallace, and Colin Raffel. 2022. Deduplicating training data mitigates privacy risks in language models. In International Conference on Machine Learning, pages 10697-10707.
- PMLR.
- Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. Qasc: A dataset for question answering via sentence compo- sition. In Proceedings of the AAAI Conference on Artificial Intelligence, 05, pages 8082-8090.
- Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885-1894. PMLR.
- Pang Wei W Koh, Kai-Siang Ang, Hubert Teo, and Percy S Liang. 2019. On the accuracy of influence functions for measuring group effects. Advances in neural information processing systems, 32.
- Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453- 466.
- Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thir- teenth international conference on the principles of knowledge representation and reasoning.
- Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct elec- tricity? a new dataset for open book question answer- ing. In EMNLP.
- Elisa Nguyen, Minjoon Seo, and Seong Joon Oh. 2023. A bayesian approach to analysing training data attri- bution in deep learning. In Thirty-seventh Confer- ence on Neural Information Processing Systems.
- Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. 2020. Estimating training data influ- ence by tracing gradient descent. Advances in Neural Information Processing Systems, 33:19920-19930.
- Melissa Roemmele, Cosmin Adrian Bejan, and An- drew S Gordon. 2011. Choice of plausible alter- natives: An evaluation of commonsense causal rea- soning. In 2011 AAAI Spring Symposium Series.
- Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. 2021. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075-18086.
Cited by
Sorted by citations
Large language models provide unsafe answers to patient-posed medical questions
Draelos, R. L., Afreen, S., Blasko, B., Brazile, T. L., Chase, N., Desai, D. P., Evert, J., Gardner, H. L., Herrmann, L., House, A. V., et al. (2026) · npj Digital Medicine · doi:10.1038/s41746-026-02428-5 · Cited by 42 · see citing works
ClassEval-T: Evaluating Large Language Models in Class-Level Code Translation
Xue, P., Wu, L., Yang, Z., Wang, C., Li, X., Zhang, Y., Li, J., Jin, R., Pei, Y., Shen, Z., et al. (2025) · Proceedings of the ACM on software engineering. · doi:10.1145/3728940 · Cited by 15 · see citing works
Can foundation language models predict fluid dynamics?
Wang, G., & Cheng, S. L. (2025) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2025.111427 · Cited by 8 · see citing works
İspir, N. G., Aslan, G., & Akay, G. (2026) · Diagnostics · doi:10.3390/diagnostics16203430 · Cited by 1 · see citing works
Park, S., & Jang, B. (2025) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3746252.3760823
- Authors
- Zhao, Y., Du, L., Ding, X., Xiong, K., Sun, Z., Jun, S., Liu, T., & Qin, B.
- Affiliations
- Harbin Institute of Technology (CN)
- Venue
- Findings of the Association for Computational Linguistics ACL 2024
- Published
- 2024
- Pages
- 9386-9406
- OpenAlex
- W4402670181
- Cited by
- 5
- References
- 0
- Citations by year
- 2026: 2 · 2025: 3
- FWCI
- 1.48
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Findings of the Association for Computational Linguistics: ACL (Findings of ACL) (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- large language models, pre-training data, machine unlearning, model capability
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
