RLHF-And-Friends/Humans-vs-Llama-Base-PPO
收藏Hugging Face2025-04-08 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/RLHF-And-Friends/Humans-vs-Llama-Base-PPO
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含了来自RLHF-And-Friends/tldr-sft测试分割的人类完成数据和RLHF-And-Friends/TLDR-Llama-3.1-8B-Base-PPO模型的完成数据。数据集中的(prompt)列包含了同时给人类和模型使用的提示。使用的是TLDR-Llama-3.1-8B-Base-PPO模型。原始数据集名为RLHF-And-Friends/tldr-sft。基于gpt-4o-mini的意见，TLDR-Llama-3.1-8B-Base-PPO模型相对于人类的胜率为0.88。
This dataset contains human completions from the RLHF-And-Friends/tldr-sft test split and completions from the RLHF-And-Friends/TLDR-Llama-3.1-8B-Base-PPO model. The column prompt includes the prompts given to both humans and the models. The model used is TLDR-Llama-3.1-8B-Base-PPO. The original dataset with prompts and human completions is named RLHF-And-Friends/tldr-sft. The winrate of TLDR-Llama-3.1-8B-Base-PPO over humans, based on the gpt-4o-mini opinion, is 0.88.
提供机构：
RLHF-And-Friends
