NCERT K12 Math - LBT DPO Pairs

A high-quality dataset of mathematical reasoning preference pairs derived from the NCERT K-12 mathematics curriculum for Direct Preference Optimization (DPO) and Learning by Teaching (LBT) training.

About Dataset

The NCERT K12 Math - LBT DPO Pairs dataset is a mathematical reasoning preference-learning dataset developed from the NCERT K-12 mathematics curriculum. It contains 44,565 examples spanning 30 mathematics chapters. Each example consists of a mathematical problem, ground-truth answer, preferred (chosen) and rejected reasoning paths, extracted final answers, multiple quality scores generated by Learning by Teaching (LBT), Teacher, and Student models, preference score differences, selection metadata, and precomputed reference model log probabilities for efficient Direct Preference Optimization (DPO) training. The citation of the dataset is - Ashwin K M. (2025). *NCERT K12 Math - LBT DPO Pairs with Precomputed Reference Logprobs*. Hugging Face Dataset. Available at: https://huggingface.co/datasets/AshwinKM2005/NCERTK12Math-LBT_DPO_Pairs. This dataset was identified and facilitated for onboarding as part of the Dataset Onboarding Support Team (DOST) initiative led by CivicDataLab (CDL), partnering with the Gates Foundation in collaboration with BHASHINI. CivicDataLab provided support for dataset discovery, validation, metadata preparation and onboarding facilitation. All dataset ownership and intellectual property rights remain with the original author(s).

Purpose of Dataset

The purpose of this dataset is to support Direct Preference Optimization (DPO) and Learning by Teaching (LBT) for mathematical reasoning using the NCERT K-12 mathematics curriculum. It enables supervised fine-tuning and alignment of large language models through preference learning and can further be used for training mathematical reasoning models, educational tutoring systems, curriculum-aware reasoning, reasoning quality evaluation, reward model development, preference optimization research, chain-of-thought reasoning, benchmark creation, and educational AI applications.

Dataset Metadata

License

Attribution 4.0 International (CC BY- 4.0)

Geographical coverage

India

Sector

Education and Skill Development

Author

Ashwin KM

Source Organisation

Digital India BHASHINI Division

Uploaded by

Nikil Augustine

Data Quality Score (Beta)

Dataset type

Structured

Frequency

Static

Time Granularity

Static

Year range

N.A.

Date & Time

24/08/26 12:01:49

Visibility

Open

Hosted / Redirected

Hosted

Data Type

Secondary

Data Collection Method

The Dataset Was Created Using The Learning By Teaching (Lbt) Framework. Mathematical Problems Were Sourced From The Ncert K-12 Mathematics Curriculum, And Multiple Reasoning Paths Were Generated For Each Problem. These Reasoning Paths Were Evaluated Using Lbt, Teacher, And Student Model Scores, After Which Preferred And Rejected Reasoning Pairs Were Selected. Each Example Contains The Mathematical Problem, Ground-truth Answer, Chosen And Rejected Rationales, Extracted Answers, Quality Scores, Selection Metadata, And Precomputed Reference Model Log Probabilities.