Idea
An open-source AI model providing step-by-step math reasoning tailored for JEE exam preparation and related benchmarks.
Research Paper
Core Innovation
This paper introduces Aryabhata 1.0, a 7B parameter model specifically fine-tuned for JEE math reasoning using curriculum learning on verified chain-of-thought data. It uniquely combines supervised fine-tuning with reinforcement learning and novel exploration strategies to improve reasoning accuracy. This approach outperforms existing models on JEE and other math benchmarks while providing interpretable stepwise solutions.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: large Indian and global exam preparation market for STEM subjects and AI-driven tutoring tools.
Potential Customers & Pain Points
- JEE Aspirants Needing Accurate Stepwise Math Solutions
- Coaching Institutes Seeking Scalable Exam Preparation Tools
- EdTech Platforms Lacking Specialized Math Reasoning Models
Business Model
Open-source foundation model with premium API access and licensing for coaching institutes and EdTech platforms.
Competitive Landscape
- OpenAI GPT-4
- Google Bard
- BYJU'S AI Tutor
Implementation Challenges
- Data Quality and Verification for Chain-of-Thought Traces
- Adoption by Traditional Coaching Institutes
- Competition from Established AI Tutoring Platforms
Validation Strategy
- Pilot integration with select JEE coaching centers
- Benchmark performance against existing AI tutoring tools
- Collect user feedback on stepwise reasoning clarity and accuracy
Research Paper Overview
Aryabhata: An exam-focused language model for JEE Math
Summary
Aryabhata 1.0 is a 7B parameter math reasoning model optimized for the Indian Joint Entrance Examination (JEE). It merges open-weight reasoning models and uses supervised fine-tuning with curriculum learning on verified chain-of-thought traces. Reinforcement learning with verifiable rewards and novel exploration strategies further boost performance. Aryabhata outperforms existing models on JEE and other math benchmarks, providing step-by-step reasoning and is released as an open-source foundation model for exam-centric education.