MathMentorDB Data Use Agreement. By requesting access you agree to the following terms, which supplement the CC BY-NC-SA 4.0 license: (1) you will use the data for non-commercial research purposes only; (2) you will not attempt to identify, de-anonymize, or contact any individual represented in the data; (3) you will not redistribute the raw data (point others to this repository instead); (4) any models or publications derived from the data must not expose personally identifying information; (5) you will report any privacy concerns you discover to the dataset maintainers via the community tab so affected content can be removed.

MathMentorDB (labeled subset)

MathMentorDB is a large-scale dataset of authentic, multi-party mathematics tutoring dialogues collected from a public online mathematics community. The full corpus comprises 5.4M messages across 200,332 conversations and 43,249 pseudonymized users. This repository releases the labeled subset described in the accompanying paper (under review; citation withheld for anonymity):

7,000 conversations / 165,275 messages, every message labeled with a 24-move hierarchical discourse taxonomy (classifier: Gemini 2.5 Flash, two-pass hierarchical classification)
Conversation-level resolution labels (resolved / abandoned / unclear) assigned by an independent classifier reading only the raw text
Canonical splits used in the paper's analyses

The full 200K-conversation corpus will be released under the same gated DUA mechanism upon publication.

Files and splits

File	Conversations	Messages	Tutor tier
`data/expert_train.jsonl`	2,500	59,124	high-activity (top 50 by composite activity score)
`data/expert_heldout.jsonl`	1,000	24,228	high-activity, held out
`data/nonexpert_train.jsonl`	2,500	58,444	occasional (rank 51+)
`data/nonexpert_heldout.jsonl`	1,000	23,479	occasional, held out

Tier note: the high-activity / occasional split reflects participation volume, not verified tutoring skill.

Schema

One conversation per JSONL line:

{
  "conversation_id": 393,
  "tutor_id": "user_00001",     // opaque pseudonym, consistent across splits
  "tutor_tier": "expert",       // activity tier (see note above)
  "tutor_score": 0.66,
  "student_id": "user_00002",   // opaque pseudonym
  "resolution": "resolved",      // resolved | abandoned | unclear
  "resolution_confidence": "high",
  "n_messages": 8,
  "messages": [
    {
      "idx": 0,
      "role": "student",
      "category": "Student-Academic",
      "move": "Knowledge-Gap",
      "category_confidence": "high",   // high | medium | low
      "move_confidence": "high",
      "message_text": "..."            // emails, platform tags, and user
    }                                   // mentions redacted
  ]
}

24-move discourse taxonomy

Four macro-categories, 24 moves, no residual "Other" class:

Tutor-Academic (7): Knowledge-Check, Guidance-Direct, Guidance-Scaffolding, Tutor-Question, Correction, Confirm-Positive, Confirm-Negative
Student-Academic (9): Knowledge-Gap, Knowledge-Recall, Inference-Attempt, Inference-Understanding, Breakthrough, Explain-Problem, Explain-Reasoning, Student-Question, Student-Confirm
Socio-Emotional (4): Encouragement, Frustration, Empathy-Rapport, Confidence-Express
Non-Academic (4): Greeting-Closing, Platform-Command, Small-Talk, Acknowledgment

Taxonomy provenance, construction process, and reliability evidence are documented in the paper appendix. Classification prompts are included in the paper's appendix; the full per-category prompt set ships with the code release.

Loading

from datasets import load_dataset
ds = load_dataset("mathmentordb/MathMentorDB",
                  data_files={"expert_train": "data/expert_train.jsonl",
                              "expert_heldout": "data/expert_heldout.jsonl",
                              "nonexpert_train": "data/nonexpert_train.jsonl",
                              "nonexpert_heldout": "data/nonexpert_heldout.jsonl"})

Provenance and processing

Conversations were extracted from public help channels of a large online mathematics community (chat exports, January 2023), segmented using platform close markers (two-annotator segmentation check, kappa = 0.94), and pseudonymized. Move and resolution labels were produced by independent LLM classifiers; run-to-run consistency on a 30-conversation held-out set is 95.0% (move level) and 98.1% (category level).

Ethics and privacy

Usernames are replaced with pseudonyms; the release excludes images and attachments. Collection and release were reviewed by the authors' institutional review board (exempt determination; details withheld for anonymous review). The data are shared under CC BY-NC-SA 4.0 plus the Data Use Agreement above, which prohibits de-anonymization and redistribution. If you find content that should be removed (personal information, identifying details), open a discussion in the community tab and it will be removed.

Citation

Paper under review; citation will be added upon publication. Until then, please cite this repository.

Downloads last month: 13

Total file size:

35.2 MB