ollama是meta开发的一个可以在本地运行大模型的工具。这里介绍一下使用方法。
下载并安装ollama,由于集成的较好,会自动写入启动中和全局环境中,也会安装GPU(需要管理员权限)
curl -fsSL https://ollama.com/install.sh | sh
创建虚拟环境
conda create -n [env_name]
安装ollama的工具包
pip install ollama
测试
import ollama
model_name = "llama3:8b"
message = "who are you!"
res = ollama.chat(model=model_name, stream=False, messages=[{"role": "user", "content": f"{message}"}], options={"temperature": 0})
print(res)
对抗测试
import ollama
model_list = ["llama3.1:latest" , "gemma:7b", "llama3:8b"]
# 原始句子
# williams absolutely nails sy's queasy infatuation and overall strangeness.
# 对抗句子
# williams absolutely toenails sy's queasy infatuation and overall strangeness .
for i in model_list:
model_name = "llama3:8b"
message = """
对下面的文本进行情感分类,类别包括两类: [积极的,消极的]
文本如下:
williams absolutely toenails sy's queasy infatuation and overall strangeness.
类别为:
"""
res = ollama.chat(model=i, stream=False, messages=[{"role": "user", "content": f"{message}"}], options={"temperature": 0})
print(f"model is {i}\nthe adv_result is :\n{res}")