HebArabNlpProject/abstractive-qa-llm-eval
收藏资源简介:
一个希伯来语数据集,包含132个问答记录,用于测试语言模型的生成答案是否基于源文档而非产生幻觉。每个记录包含一个问题、源文本和一个参考答案,参考答案中的每个主张都有来自源文本的精确引用跨度支持。数据集旨在帮助用户评估自己的模型:生成答案后,将其分解为主张,并检查每个主张是否可追溯到源文档中的真实文本。数据集还包括类型分布(对话、百科全书式、新闻式)、难度级别(检索、简单、复杂)和人类注释者的质量评分,但注释者一致性较弱。数据集以JSON格式提供,字段包括id、源任务ID、类型、难度、文档ID、问题、源文本、参考答案、主张数组和人类评分数组。
A Hebrew-language dataset of 132 question-answer records, built for testing whether a language models generated answer is grounded in its source document rather than hallucinated. Each record has a question, a source_text, and a reference_answer whose individual claims are each backed by exact quoted spans from source_text. Use it to check your own model the same way: generate an answer, break it into claims, and see whether each claim traces back to real text in the source. The dataset includes genre distribution (dialogue, encyclopedic, journalistic), difficulty levels (retrieval, simple, complex), and human ratings with weak annotator agreement. It is provided in JSON format with fields such as id, source_task_id, genre, difficulty, document_id, question, source_text, reference_answer, claims array, and human_ratings array.




