[Dataset] MathDial Dataset

[jadongeathers]

[Dataset] MathDial Corpus (math word problem tutoring dialogs)

Dataset name

MathDial (ConvoKit format)

Brief description

A collection of human–tutor dialogues focused on solving math word problems.
This dataset is adapted from the original MathDial corpus and converted into the ConvoKit format.
It contains both the original 4-way tutor intents and fine-grained 11-way tutor intents (for teacher turns).

---

Dataset details

Speaker-level

Speakers are either teachers or students. Some conversations include named students (e.g., Cody, Mariana), while others use the generic “Student.”
Each speaker is represented consistently within a conversation.

• id: <Name>_<conversation_id> (e.g., Teacher_conv12, DeAndre_conv284)
• meta.role: normalized role — "teacher" or "student"
• meta.role_raw: the original value from the TSV (e.g., "Teacher", "Student", "DeAndre")
• meta.conversation_id: conversation this speaker belongs to (e.g., conv284)
• meta.split: dataset split (train, val, test)

Utterance-level

Each conversational turn is an utterance.

• id: global utterance identifier
• speaker: speaker who produced the utterance
• conversation_id: conversation identifier (e.g., conv0, conv1, …)
• reply_to: previous utterance in the thread (None if start)
• timestamp: not available
• text: textual content of the utterance

Metadata for each utterance includes:

• intent_4: coarse 4-way intent label
• intent_11: fine 11-way intent label (teacher turns only)
• qid: problem identifier

Conversation-level

Metadata for each conversation includes:

• conversation_id: global string identifier (conv0, conv1, …)
• split: which split (train, val, test)
• qid: problem ID
• scenario: description of the problem context
• question: math problem text
• ground_truth: correct solution
• student_incorrect_solution: incorrect solution given by the student
• student_profile: information about the student (if available)
• teacher_described_confusion, self-correctness, self-typical-confusion, self-typical-interactions: additional annotation fields

Corpus-level

• name: mathdial
• source_repo: https://github.com/Kpetyxova/autoTree/tree/main/mathdial
• description: MathDial dialogs converted to ConvoKit with both 4-way and 11-way tutor intents. IDs are namespaced by split.
• num_utterances: 8,466
• num_conversations: 521
• num_speakers: 1,043

---

Licensing information

The MathDial dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

---

Citation

Petukhova, Kseniia, and Ekaterina Kochmar. "Intent Matters: Enhancing AI Tutoring with Fine-Grained Pedagogical Intent Annotation." arXiv preprint arXiv:2506.07626, 2025.

---

Contact

Jadon Geathers — jag569@cornell.edu

---

Access

Here is a link to the zipped corpus:
mathdial.zip

---

Statistics

• Conversations: 521
• Utterances: 8,466
• Speakers: 1,043

Top fine-grained (11-way) tutor intents:

Intent	Count
Revealing Strategy	1141
Revealing Answer	895
Guiding Student Focus	687
Seek Strategy	658
Asking for Explanation	653
Seeking Self Correction	643
Seeking World Knowledge	257
Greeting/Farewell	217
Recall Relevant Information	93
Perturbing the Question	89
General Inquiry	40

---

Example usage

from convokit import Corpus

corpus = Corpus("PATH_TO/mathdial")
corpus.print_summary_stats()