oumi-ai/MiniMath-R1-1.5B

Supervised fine-tune of DeepSeek-R1-Distill-Qwen-1.5B using oumi-ai/MetaMathQA-R1.

Achieves 44.4% accuracy on MMLU-Pro-Math, the highest of any model with <=1.5B parameters.

Improves the base model's accuracy by +6 points.

• Developed by: Oumi AI
• Model type: Small Language Model
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: DeepSeek-R1-Distill-Qwen-1.5B
• Demo: Fine-Tuning Notebook

Uses

Use as a conversational assistant for solving math problems with an exposed thought process.

Out-of-Scope Use

Smaller LLMs have limited capabilities and should be used with caution. Avoid using this model for purposes outside of mathematics.

Bias, Risks, and Limitations

This model was finetuned with DeepSeek-R1 data on top of an R1-distill model, so any biases or risks associated with those models may be present.

Training Details

Training Data

Training data: oumi-ai/MetaMathQA-R1

Training Procedure

Training notebook: Fine-Tuning Notebook

Evaluation

Environmental Impact

• Hardware Type: H100
• Hours used: 0.8 (0.1 * 8 GPUs)
• Cloud Provider: Google Cloud Platform
• Compute Region: us-east5
• Carbon Emitted: 0.07 kg

Citation

@misc{miniMathR1_2025,
  author = {Jeremiah Greer},
  title = {MiniMath-R1-1.5B},
  month = {February},
  year = {2025},
  url = {https://huggingface.co/oumi-ai/MiniMath-R1-1.5B}
}

@software{oumi2025,
  author = {Oumi Community},
  title = {Oumi: an Open, End-to-end Platform for Building Large Foundation Models},
  month = {January},
  year = {2025},
  url = {https://github.com/oumi-ai/oumi}
}