ContinuousBench/Geminon

ContinuousBench — Geminon (2025_09)是一个由ContinuousBenchCuration管道生成的数据集，主要用于问答（QA）任务。数据集包含多种配置：index（公共和敏感分割）、corpus_large（完整去重语料库，含1,523,754篇文章，平均118个token）、corpus_medium（平衡的100万篇文章子样本，平均116个token）、corpus_small（平衡的20万篇文章子样本，平均112个token）、qa_small和qa_medium（分别基于corpus_small和corpus_medium的问答数据，包含公共和敏感验证/测试分割）。数据集涵盖五种文章类型：wiki、sensitive_wiki、journal、comparison和evolution，涉及多个属性（如能力、攻击、防御等）的问答支持。统计信息使用Gemma 3 tokenizer计算，并提供每个属性的支持计数，公共Geminons有多个支持文章，敏感Geminons仅有一个支持文章。ContinuousBench — Geminon (2025_09) is a dataset generated by the ContinuousBenchCuration pipeline, primarily designed for question answering (QA) tasks. The dataset includes multiple configurations: index (with public and sensitive splits), corpus_large (a full deduplicated corpus containing 1,523,754 articles with an average of 118 tokens), corpus_medium (a balanced subsample of 1,000,000 articles with an average of 116 tokens), corpus_small (a balanced subsample of 200,000 articles with an average of 112 tokens), as well as qa_small and qa_medium (QA datasets based on corpus_small and corpus_medium respectively, featuring public and sensitive validation/test splits). The dataset covers five article types: wiki, sensitive_wiki, journal, comparison, and evolution, and supports QA for multiple attributes including capabilities, attacks, defenses, and others. Tokenization statistics are calculated using the Gemma 3 tokenizer, and support counts are provided for each attribute. Public Geminons have multiple supporting articles, while sensitive Geminons only have one supporting article.