Datasets:

lasha-nlp
/

NYT_Memorization

The 📰NYT datasets serve as a benchmark designed to evaluate methods to identify memorized training data or to infer membership, specifically from OpenAI models that are released before 2023.

🔧 Memorized Training Data

This dataset contains examples provided in Exhibit-J of the New York Times Lawsuit (label=1).

Snippet is the article.
Prefix is the prefix provided to the model according to the evidence in Exhibit-J.
Completion is the original article content that follows the prefix.

🔧 Non-Member Data

This dataset also contains excerpts of CNN articles scraped in 2023 (label=0).

📌 Applicability

The datasets can be applied to OpenAI models released before 2023.

Loading the datasets

To load the dataset:

from datasets import load_dataset

dataset = load_dataset("lasha-nlp/NYT_Memorization")

Label 0: Refers to the CNN data that cannot have been memorized. Label 1: Refers to text data from NYT articles.

🛠️ Codebase

For evaluating information-guided probes, visit our GitHub repository.

⭐ Citing our Work

If you find our code or data useful, kindly cite our work:

@misc{ravichander2025informationguidedidentificationtrainingdata,
      title={Information-Guided Identification of Training Data Imprint in (Proprietary) Large Language Models}, 
      author={Abhilasha Ravichander and Jillian Fisher and Taylor Sorensen and Ximing Lu and Yuchen Lin and Maria Antoniak and Niloofar Mireshghallah and Chandra Bhagavatula and Yejin Choi},
      year={2025},
      eprint={2503.12072},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.12072}, 
}