TIGER-Lab/MAmmoTH-Coder-34B

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Sep 11, 2023License:mitArchitecture:Transformer0.0K Open Weights Cold

MAmmoTH-Coder-34B is a 34 billion parameter language model developed by TIGER-Lab, based on Code Llama, and specifically fine-tuned for general math problem-solving. It utilizes a hybrid instruction tuning approach with Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales from the MathInstruct dataset. This model excels at generating comprehensive solutions for diverse mathematical problems, making it suitable for applications requiring robust mathematical reasoning.

Loading preview...

MAmmoTH-Coder-34B: A Math Generalist Model

MAmmoTH-Coder-34B is a 34 billion parameter model from the MAmmoTH series, developed by TIGER-Lab, and built upon the Code Llama architecture. It is specifically designed for general math problem-solving, leveraging a unique hybrid instruction tuning approach.

Key Capabilities

  • Hybrid Reasoning: Trained on the MathInstruct dataset, it effectively combines Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales to solve mathematical problems. This allows for both step-by-step textual explanations and executable code-based solutions.
  • Diverse Math Coverage: The training dataset, MathInstruct, is compiled from 13 math rationale datasets, ensuring extensive coverage across various mathematical fields.
  • Strong Performance: Evaluation results demonstrate its proficiency across multiple math benchmarks, including GSM, MATH, AQuA, and MMLU, particularly when using the hybrid decoding strategy.

When to Use This Model

  • Mathematical Problem Solving: Ideal for applications requiring robust solutions to general math problems.
  • Educational Software: Can be integrated into tutoring systems or platforms that need to generate detailed math solutions.
  • Research in Mathematical Reasoning: Useful for researchers exploring advanced reasoning techniques in LLMs, especially the interplay between CoT and PoT.

Limitations

While designed as a math generalist, its performance may vary with problem complexity and specific mathematical domains. It may not comprehensively cover all highly specialized mathematical fields.