Ollama 实测：Llama 3 的审查比 Llama 2 少很多
Llama 3 is not very censored · Ollama Blog
Ollama 在博客中称 Llama 3 的误拒率不到 Llama 2 的三分之一，可讨论的话题范围更广。文中用同一组提示词对比了 Llama 3 8B 与 Llama 2 7B 的输出，两者均采用 4-bit 整数量化。用户可下载 Ollama 并运行 llama3 在本地体验。
推荐理由
精选AI 评分 61
Ollama 用同一组提示词对比 Llama 3 8B 与 Llama 2 7B 的拒答表现，可直观看到新模型在话题覆盖上的变化。
主题
正文
当前显示原文，中文全文译文尚未就绪。
April 19, 2024
Llama 3 feels significantly less censored than its predecessor.
The Llama 3 models have substantially lower false refusal rates, with less than 1⁄3 the amount of false refusals when compared to Llama 2, making it possible to discuss a wider range of interesting topics!
Examples
Below are examples comparing results for series of prompts between Llama 3 8B and Llama 2 7B, both optimized using to 4-bit integer quantization:
To run Llama 3 locally, Download Ollama and run llama3:
Sources