Aryabhata 1.0 is a 7 billion-parameter small mathematical inference model optimized for the Indian entrance exam, JEE. While existing large-scale language models (LLMs) are often inadequate for training, Aryabhata 1.0 combines powerful open-weighted inference models and was developed through supervised learning fine-tuning (SFT) and curriculum learning using proven course of thought (CoT) tracking. It further improves performance by applying novel exploration strategies, such as Reinforcement Learning with Verifiable Rewards (RLVR) using the A2C objective and group relative advantage estimation, as well as adaptive group sizing and temperature control. It outperforms existing models in accuracy and efficiency on in-distribution benchmarks such as JEE Main 2025 and out-of-distribution benchmarks such as MATH and GSM8K, and provides educationally useful step-by-step inference. Aryabhata 1.0 is released as a foundational model for developing open-source, test-focused small language models.
