Overview

kevinpro/MistralMathOctopus-7B is a 7 billion parameter language model developed by kevinpro, designed to excel in multilingual reasoning and mathematical problem-solving. This model is an enhanced version of MistralMathOctopus 7B, further optimized using the Multilingual Alignment-as-Preference Optimization (MAPO) technique. The research behind this model is detailed in the paper "MAPO: Advancing Multilingual Reasoning through Multilingual Alignment-as-Preference Optimization" arXiv:2401.06838.

Key Capabilities

• Superior Multilingual Math Reasoning: Achieves high scores on challenging multilingual mathematical benchmarks.
• MAPO Optimization: Utilizes a novel alignment-as-preference optimization method to boost performance.
• Competitive Benchmarks: Outperforms several other 7B and even some 13B models, including GPT-3.5-Turbo, MAmmoTH, WizardMath, MetaMath, and QAlign, on datasets like MSVAMP, MGSM, and MNumGLUESub.
  • MSVAMP: Scores 74.6, significantly higher than GPT-3.5-Turbo's 46.6.
  • MGSM: Achieves 67.3, surpassing GPT-3.5-Turbo's 42.2.
  • MNumGLUESub: Reaches 70.0, compared to GPT-3.5-Turbo's 49.4.

Good for

• Applications requiring advanced mathematical reasoning in a multilingual context.
• Research and development in multilingual NLP and mathematical problem-solving.
• Use cases where strong performance on complex reasoning tasks is critical.