latkes/inside-out-replication-v2-external-scores

该数据集为Inside-Out V2提供每个（问题，答案）对的外部分数及法官标签。覆盖了P(a|q)（给定问题的答案概率）、P_norm(a|q)（长度归一化概率）、P(True)以及两种验证提示变体。用于计算外部K/K*指标。数据集包含1,752,198行和20列，列包括问题ID、答案文本、法官标签（CORRECT或INCORRECT）、对数概率、概率值、归一化概率、P(True)分数、全词汇概率质量、残差质量，以及不同验证提示变体的分数和残差。还包含主题模型、Wikidata关系和数据集分割（dev或test）信息。This dataset provides external scores and judge labels for each (question, answer) pair for Inside-Out V2. It covers P(a|q) (probability of the answer given the question), P_norm(a|q) (length-normalized probability), P(True), and two validation prompt variants, and is used to calculate the external K/K* metrics. The dataset consists of 1,752,198 rows and 20 columns, with columns including question ID, answer text, judge label (either CORRECT or INCORRECT), log probability, probability value, normalized probability, P(True) score, full vocabulary probability mass, residual mass, as well as scores and residuals for different validation prompt variants. It also includes information on topic models, Wikidata relationships, and dataset splits (dev or test).