kevinpro/MathOctopus-MAPO-DPO-7B

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Feb 14, 2024License:apache-2.0Architecture:Transformer Open Weights Cold

kevinpro/MathOctopus-MAPO-DPO-7B is a 7 billion parameter language model developed by Shuaijie She et al., fine-tuned using Multilingual Alignment-as-Preference Optimization (MAPO-DPO). This model specializes in multilingual reasoning tasks, particularly mathematical problem-solving across various languages. It demonstrates strong performance on benchmarks like MSVAMP, MGSM, and MNumGLUESub, often outperforming other 7B models and even GPT-3.5-Turbo in specific multilingual math reasoning scenarios.

Loading preview...

MathOctopus-MAPO-DPO-7B: Multilingual Reasoning with Preference Optimization

kevinpro/MathOctopus-MAPO-DPO-7B is a 7 billion parameter model developed by Shuaijie She et al., leveraging a novel approach called Multilingual Alignment-as-Preference Optimization (MAPO-DPO). This fine-tuning method is designed to enhance the model's capabilities in multilingual reasoning, with a particular focus on mathematical problem-solving.

Key Capabilities & Performance

  • Multilingual Mathematical Reasoning: The model is specifically optimized for complex mathematical tasks presented in multiple languages.
  • Benchmark Excellence: It achieves competitive results on challenging multilingual math benchmarks:
    • MSVAMP: Scores 57.4, surpassing GPT-3.5-Turbo (46.6) and other 7B models like MetaMath 7B (46.2).
    • MGSM: Achieves 41.6, comparable to GPT-3.5-Turbo (42.2).
    • MNumGLUESub: Scores 50.4, outperforming GPT-3.5-Turbo (49.4).
  • DPO Fine-tuning: The use of Direct Preference Optimization (DPO) with multilingual alignment data contributes to its strong performance in reasoning tasks.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Multilingual Math Solvers: Developing tools that can understand and solve mathematical problems presented in various languages.
  • Educational AI: Creating intelligent tutoring systems or assessment tools for math in diverse linguistic contexts.
  • Cross-lingual Reasoning: Any task demanding robust logical and mathematical reasoning across different languages.