Nemotron-MIND is a dataset released by NVIDIA in 2024, containing over 138 billion tokens of structured mathematical conversations, with 231.6 million records and a total storage size of 827 GB. It covers seven different social conversation styles, including debate, problem solving, and teacher-student dialogue, and is associated with the research paper "MIND: Math Informed syNthetic Dialogues for Pretraining LLMs", aimed at enhancing the mathematical reasoning capabilities of large language models through synthetic conversational data.

Dataset Composition

• TWO STUDENTS: Conversations between two students.
• TEACHER STUDENT: Conversations between a teacher and a student.
• TWO PROFESSORS: Conversations between two professors.
• DEBATE: Conversations in a debate format.
• PROBLEM SOLVING: Conversations centered around problem analysis and resolution.
• LAYMAN KNOWALL: Conversations in the style of an omniscient layperson.
• INTERVIEW: Conversations in an interview format.

Citation

@inproceedings{
akter2025mind,
title={{MIND}: Math Informed syNthetic Dialogues for Pretraining {LLM}s},
author={Syeda Nahida Akter and Shrimai Prabhumoye and John Kamalu and Sanjeev Satheesh and Eric Nyberg and Mostofa Patwary and Mohammad Shoeybi and Bryan Catanzaro},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=TuOTSAiHDn}
}