MiLM-6B is a large-scale pretrained language model developed by Xiaomi, with a parameter scale of 6.4 billion. According to official materials, the model achieves the best results among models of the same size on both C-Eval and CMMLU.
In terms of purpose, MiLM-6B is a Chinese pretrained language model, mainly used for natural language processing-related tasks, and its performance evaluation focuses on Chinese knowledge and reasoning abilities.
Regarding its main capabilities, official materials only state that its parameter scale is 6.4 billion and that it achieves the best results among models of the same size on the two Chinese evaluation benchmarks C-Eval and CMMLU. Among them, the C-Eval dataset is a comprehensive Chinese foundation model evaluation dataset covering 52 disciplines and four difficulty levels; the CMMLU dataset is a comprehensive Chinese evaluation benchmark specifically used to evaluate language models' knowledge and reasoning abilities in Chinese contexts, covering 67 topics ranging from basic disciplines to advanced professional levels. The materials do not provide other specific capabilities, functions, or usage methods.
In terms of usage, the model is provided through a GitHub repository, and the repository address is https://github.com/XiaoMi/MiLM-6B , and the page indicates that one can participate in development by creating a GitHub account. The materials do not provide installation commands, invocation code, license, model weight download methods, or running steps, so this information is not provided for now.
In terms of applicable scenarios, based on the available materials, it can be inferred to be suitable for knowledge question answering and reasoning evaluation or research purposes in Chinese contexts, but the materials do not clearly list specific application scenarios, so no further details are provided for now.
In terms of limitations, the materials only state that the parameter scale is 6.4 billion and do not provide other technical limitations, hardware requirements, language support scope, or usage constraints, so they are not provided for now.
