YU-MO/Yumo-nano
Yumo Nano is a 1.5 billion parameter mathematics-specialized language model developed by OpceanAI, fine-tuned from DeepScaleR-1.5B-Preview. It features a 2,048 token context length and was trained using a three-phase curriculum on a consumer RTX 4080. This model excels in mathematical reasoning, outperforming its base model across all five evaluated benchmarks, including significant gains on OlympiadBench and Minerva Math.
Loading preview...
Yumo Nano: A 1.5B Math Specialist
Yumo Nano, developed by OpceanAI, is a 1.5 billion parameter language model specifically fine-tuned for mathematics. Derived from DeepScaleR-1.5B-Preview, it represents the first release in the Yumo model family. This model was trained on a consumer RTX 4080 using a unique three-phase supervised fine-tuning curriculum designed to establish a consistent mathematical personality, deepen domain-specific capabilities, and then consolidate them.
Key Capabilities
- Superior Mathematical Reasoning: Yumo Nano surpasses its base model, DeepScaleR-1.5B, across all five evaluated benchmarks, including AIME 2024, MATH 500, AMC 2023, Minerva Math, and OlympiadBench. It achieves a notable +2.9 point improvement on OlympiadBench, which is highly resistant to gains at this scale, and +2.1 points on Minerva Math, targeting multi-step scientific and mathematical reasoning.
- Defined Mathematical Identity: The model is imbued with a persistent behavioral baseline as a curious, precise, and direct mathematical AI, covering arithmetic to abstract algebra. This identity is trained into the model weights, not merely injected via system prompts.
- Efficient Training: Achieved through a three-phase curriculum on an RTX 4080, demonstrating effective fine-tuning even on consumer-grade hardware.
Good for
- Mathematical Problem Solving: Excels at a wide range of mathematical tasks, from basic arithmetic to complex proofs and competition-level reasoning.
- Step-by-Step Explanations: Designed to use clear notation and explain reasoning comprehensively.
- Resource-Constrained Environments: Its 1.5B parameter size and efficient performance make it suitable for deployment on consumer hardware like an RTX 4080.