TIGER-Lab/MAmmoTH-Coder-13B

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Sep 11, 2023License:mitArchitecture:Transformer0.0K Open Weights Cold

The TIGER-Lab/MAmmoTH-Coder-13B is a 13 billion parameter large language model developed by TIGER-Lab, specifically fine-tuned for general math problem-solving. Built upon the Code Llama base model, it excels at mathematical reasoning by leveraging a hybrid instruction tuning approach that combines Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales. This model is optimized for generating comprehensive solutions to math problems, including programmatic approaches, making it suitable for educational and problem-solving applications.

Loading preview...

MAmmoTH-Coder-13B: A Math Generalist LLM

MAmmoTH-Coder-13B is a 13 billion parameter open-source large language model from TIGER-Lab, specifically designed for general math problem-solving. It is built on the Code Llama base model and distinguishes itself through a unique hybrid instruction tuning approach.

Key Capabilities

  • Hybrid Reasoning: Utilizes both Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales for robust mathematical problem-solving.
  • Specialized Training Data: Fine-tuned on the meticulously curated MathInstruct Dataset, which comprises 13 diverse math rationale datasets.
  • Strong Math Performance: Achieves an average score of 54.9% across various math benchmarks (GSM, MATH, AQuA, NumG, SVA, Mat, Sim, SAT, MMLU) using hybrid decoding, outperforming its CoT and PoT-only counterparts.
  • Code-Based Solutions: As a 'Coder' variant, it is particularly adept at generating programmatic solutions for math problems.

Good For

  • Educational Software: Integrating into applications that require solving or explaining mathematical problems.
  • Tutoring Systems: Providing detailed, step-by-step or programmatic solutions for students.
  • Research in Mathematical Reasoning: Exploring the effectiveness of hybrid CoT and PoT approaches in LLMs.

This model is intended for research purposes and applications requiring comprehensive mathematical problem-solving capabilities, including those benefiting from code-based reasoning.