Dataset Card for MathEd-PII

Dataset Summary

MathEd-PII is a dataset focused on de-identifying Personally Identifiable Information (PII) within mathematics education and tutoring transcripts. This dataset contains surrogate ground truth data generated from question-anchored, on-demand mathematics tutoring sessions, providing a valuable resource for training and evaluating PII detection and redaction models in educational contexts.

Supported Tasks

• token-classification, named-entity-recognition: The dataset can be used to train models to identify and classify PII entities within educational dialogues.
• text-generation: Can be used for evaluating text sanitization and surrogate generation models.

Languages

The text in the dataset is primarily in English (en).

Dataset Structure

Data Instances

Each instance in the dataset represents a tutoring session transcript with labeled PII entities and their corresponding surrogate replacements. See an exerpt in the example below.

{
  "role":"volunteer",
  "content":"Hi Chloe! What can I help you with today?",
  "session_id":16592,
  "sequence_id":0,
  "annotations":[
    {
      "pii_type":"PERSON",
      "surrogate":"Chloe",
      "start":3,
      "end":8
    }
  ]
}

{
  "role":"student",
  "content":"hello!",
  "session_id":16592,
  "sequence_id":1,
  "annotations":[]
}

Data Fields

• role: The role of the speaker, "volunteer" or "student".
• content: The text content of the message.
• session_id: The ID of the tutoring session.
• sequence_id: The sequence number of the message within the session.
• annotations: A list of PII annotations, each containing:
  • pii_type: The type of PII. In total, there are 14 types of PII (number of instances in parentheses): PERSON (1,424), URL (187), LOCATION (121), GRADE_LEVEL (107), SCHOOL (73), COURSE_NUMBER (40), NRP (Nationality, Religious or Political group;25), AGE (8), DATE (4), US_DRIVER_LICENSE (2), PHONE_NUMBER (2), and IP_ADDRESS (1).
  • surrogate: The surrogate replacement for the PII.
  • start: The starting index of the PII in the content.
  • end: The ending index of the PII in the content.

Dataset Creation

Curation Rationale

This dataset was created to address the lack of specialized, open-access datasets for PII de-identification in educational domains, specifically online tutoring. It enables researchers to build safer, privacy-preserving AI tools for education.

Source Data

The original source data comes from math tutoring transcripts collected from a U.S.-based online tutoring platform.

Annotations

The dataset includes LLM-generated annotations for PII deteaction and surrogate replacement based on the pre-redacted tutoring transcripts. Note, over-redaction was observed in the original transcripts. The LLM procedure accounted for this by human-in-the-loop evaluation. Please check the paper for more details.

Considerations for Using the Data

For Privacy Preservation

This dataset supports the development of privacy-preserving technologies in education, enabling safer sharing and analysis of tutoring data for research and AI development.

For Math Tutoring Studies

Due to some over-redaction in the original data, this dataset's ability to fully reflect real-world math tutoring processes may be slightly affected, as some mathematical content was inferred by an LLM post hoc rather than derived directly from the raw transcripts.

Additional Information

Licensing Information

The dataset is released under dual licenses:

• MIT License (typically for accompanying code/scripts)
• CC-BY 4.0 License (Creative Commons Attribution 4.0 International) for the dataset content.

Citation Information

@article{zhou2026utility,
  title={Utility-Preserving De-Identification for Math Tutoring: Investigating Numeric Ambiguity in the MathEd-PII Benchmark Dataset},
  author={Zhou, Zhuqian and Vanacore, Kirk and Ahtisham, Bakhtawar and Lee, Jinsook and Pietrzak, Doug and Hedley, Daryl and Dias, Jorge and Shaw, Chris and Sch{\"a}fer, Ruth and Kizilcec, Ren{\'e} F},
  journal={arXiv preprint arXiv:2602.16571},
  year={2026}
}