SciQAG
收藏资源简介:
SciQAG是一个由澳大利亚新南威尔士大学等机构合作开发的数据集,专注于自动生成科学文献中的问答对。该数据集包含960,000个问答对,源自96,000篇科学论文,旨在通过精细化的评估框架提升大型语言模型在科学领域的应用。SciQAG通过使用开源的大型语言模型自动从全文科学论文中提取信息,生成高质量的问答对,并通过五维评估指标确保其质量。此数据集不仅可作为评估模型在科学领域封闭式问答任务的基准,还可用于训练大规模模型,以避免模型因使用完全由语言模型自生成的训练数据而导致的崩溃。
SciQAG is a dataset co-developed by institutions including the University of New South Wales in Australia, focusing on automated generation of question-answer pairs from scientific literature. The dataset encompasses 960,000 question-answer pairs sourced from 96,000 scientific papers, and is designed to advance the application of large language models in the scientific domain via a refined evaluation framework. SciQAG generates high-quality question-answer pairs by utilizing open-source large language models to automatically extract information from full-text scientific papers, and ensures the quality of these pairs through five-dimensional evaluation metrics. This dataset can not only serve as a benchmark for evaluating models on closed-domain scientific question answering tasks, but also be employed to train large-scale models to prevent model collapse resulting from training data entirely generated by language models themselves.

- 1SciQAG: A Framework for Auto-Generated Scientific Question Answering Dataset with Fine-grained Evaluation澳大利亚新南威尔士大学 · 2024年



