Authors: Arindam Mitra, Hamed Khanpour, Corby Rosset, Ahmed Awadallah

Published on: February 16, 2024

Impact Score: 8.0

Arxiv code: Arxiv:2402.14830

Summary

• What is new: Orca-Math achieves over 86% accuracy on GSM8k, a significant improvement over other models without requiring ensembling, verifiers, or external tools.
• Why this is important: Small language models struggle with mathematical word problems, requiring massive models or complex techniques for high accuracy.
• What the research proposes: A novel 7-billion-parameter model, Orca-Math, utilizing a high-quality synthetic dataset and iterative learning from feedback.
• Results: Orca-Math surpassed both larger and smaller models in accuracy, achieving 86.81% on GSM8k without auxiliary techniques.

Technical Details

Technological frameworks used: Mistral-7B

Models used: Orca-Math

Data used: 200K high quality synthetic dataset created via a multi-agent setup

Potential Impact

Educational technology, online tutoring platforms, and AI-based problem-solving applications

Want to implement this idea in a business?

We have generated a startup concept here: MathBridge.