MetaMath 13B
About
MetaMath 13B is a specialized large language model fine-tuned for mathematical reasoning, built upon the LLaMA-7B architecture. It significantly enhances mathematical problem-solving by leveraging the MetaMathQA dataset, which includes augmented content from existing datasets like GSM8K and MATH. This model excels in benchmark tests, outperforming many open-source models of similar size, although it falls short compared to some closed-source models like GPT-3.5-Turbo. While the MetaMathQA dataset is available to the public, the details of its creation and fine-tuning are not fully transparent. MetaMath 13B's proficiency is notably strong in mathematical domains, but less so elsewhere, with performance varying based on question complexity and phrasing. Available in quantized formats like GPTQ and AWQ, it supports various inference tools, offering flexibility in resource usage.