jz666/gemma2-ultrafeedback-ppl-split-40

该数据集包含了用于文本生成的提示（prompt）和相应的生成响应，以及一系列用于评估生成响应的指标。数据集分为训练集、测试集以及多个特定的子训练集，用于不同条件下的模型训练和评估。每个样本包括提示ID、提示文本、所有生成的响应、响应的RM评分、选择的响应内容及其角色、被拒绝的响应内容及其角色，以及选择的响应的困惑度（ppl）和负对数似然（nll）。

The dataset includes prompts for text generation and corresponding generated responses, along with a series of metrics for evaluating the generated responses. The dataset is split into training sets, test sets, and several specific sub-training sets for model training and evaluation under different conditions. Each sample includes prompt ID, prompt text, all generated responses, RM scores for responses, chosen response content and role, rejected response content and role, as well as perplexity (ppl) and negative log-likelihood (nll) for the chosen response.