TIGER-Lab/MAmmoTH-70B

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Sep 11, 2023License:mitArchitecture:Transformer0.0K Open Weights Cold

MAmmoTH-70B by TIGER-Lab is a 69 billion parameter large language model, based on Llama-2, specifically fine-tuned for general math problem-solving. It excels at mathematical reasoning by leveraging a hybrid approach of Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales, trained on the MathInstruct dataset. This model is designed to provide comprehensive solutions to diverse mathematical problems, making it suitable for educational and tutoring applications.

Loading preview...

MAmmoTH-70B: A Math Generalist LLM

MAmmoTH-70B, developed by TIGER-Lab, is a 69 billion parameter large language model built upon the Llama-2 architecture. It is specifically designed and fine-tuned for general math problem-solving, distinguishing itself through a unique hybrid instruction tuning approach.

Key Capabilities

  • Hybrid Reasoning: Utilizes both Chain-of-Thought (CoT) and Program-of-Thought (PoT) rationales to solve math problems, offering comprehensive and verifiable solutions.
  • Specialized Training Data: Fine-tuned on the meticulously curated MathInstruct Dataset, which combines 13 diverse math rationale datasets, including newly curated ones.
  • Broad Mathematical Coverage: The training data ensures extensive coverage across various mathematical fields, enhancing its generalizability.
  • Strong Performance: Achieves an average score of 63.4% across multiple math benchmarks (GSM, MATH, AQuA, NumG, SVA, Mat, Sim, SAT, MMLU) when using its hybrid decoding strategy.

Intended Uses

  • Educational Software: Ideal for integration into platforms that require robust math problem-solving capabilities.
  • Tutoring Systems: Can serve as a core component for AI-powered tutoring, providing step-by-step solutions and explanations.
  • Research: A valuable tool for researchers exploring advanced mathematical reasoning in LLMs.

MAmmoTH-70B is particularly effective for tasks requiring detailed mathematical reasoning and the generation of programmatic solutions, offering a significant advantage over models trained solely on CoT.