onlyoneaman/mlx-llm-bench
收藏资源简介:
MLX LLM Bench是一个用于在统一内存有限的Apple Silicon Mac上评估本地大型语言模型(LLM)的分类和指令遵循基准测试。目标硬件为16 GB Mac mini类别。数据集包含125个样本,其中100个用于分类任务,25个用于IFEval指令遵循任务,平衡了9个主题类别,并具有明确的难度字段。分类任务包括情感分析(正面/负面)、主题分类(世界/体育/商业/技术)和垃圾邮件检测(垃圾邮件/正常邮件),而IFEval任务评估模型遵循特定指令的能力,如精确字数统计、字母排除、全大写输出、JSON模式遵循和段落计数。数据集旨在为本地LLM在资源受限环境下的性能提供标准化评估。
Classification + instruction-following benchmark for local LLMs on Apple Silicon Macs with limited unified memory. Target hardware: 16 GB Mac mini class. The dataset contains 125 examples (100 classification + 25 IFEval instruction-following), balanced across 9 topic classes, with an explicit difficulty field. It includes tasks such as sentiment analysis (positive/negative), topic classification (world/sports/business/tech), spam detection (spam/ham), and IFEval for measuring instruction-following capabilities like exact word count, letter exclusion, ALL CAPS output, JSON-schema adherence, and paragraph count. Designed to evaluate the performance of local LLMs in resource-constrained environments.





