使用的系统(base)➜ ~ lsb_release-aNo LSB modules are available. Distributor ID: Pop Description: Pop!_OS24.04LTS Release:24.04Codename: noble创建虚拟环境我们首先创建一个OpenCompass的Python虚拟环境conda create--nameopencompasspython3.10-yconda activate opencompass安装OpenCompass我们使用源码方式安装OpenCompass克隆OpenCompossgitclone https://github.com/open-compass/opencompass opencompass修改依赖的版本安装依赖cdopencompass pipinstall-e.准备数据集我们使用OpenComposs的核心数据集,把数据集放到OpenComposs项目的根目录下。wgethttps://github.com/open-compass/opencompass/releases/download/0.2.2.rc1/OpenCompassData-core-20240207.zipunzipOpenCompassData-core-20240207.zip评估下载模型我们从ModelScope下载模型安装modelscopepipinstallmodelscope下载模型我们对Qwen/Qwen3-0.6B模型进行评估下载到llm目录下。frommodelscopeimportsnapshot_download model_dirsnapshot_download(Qwen/Qwen3-0.6B,cache_dirllm)开始评估python run.py\--datasetsdemo_gsm8k_base_gen demo_math_base_gen\--hf-type chat\--hf-path /home/gillbert/code/opencompass/llm/models/Qwen--Qwen3-0.6B/snapshots/master\--debug参数说明datasets 指定数据集hf-type HuggingFace 模型类型可选值为 chat 或 basehf-path HuggingFace 模型路径执行命令输出结果如下The markdown format results is as below: | dataset | version | metric | mode | master_hf | |----- | ----- | ----- | ----- | -----| | demo_gsm8k | 17d0dc | accuracy | gen | 56.25 | | demo_math | db136b | accuracy | gen | 21.88 | 09/12 06:34:14 - OpenCompass - INFO - write markdown summary to /home/gillbert/code/opencompass/outputs/default/20260912_060759/summary/summary_20260912_060759.md参考文档https://doc.opencompass.org.cn/zh_CN/