Beemo
收藏资源简介:
Beemo数据集由托洛卡人工智能创建,包含19.6k条文本,涵盖人类编写、机器生成和专家编辑的内容。数据集内容丰富,包括创意写作、重写、总结和开放及封闭式问答等多种用例。创建过程涉及使用No Robots数据集作为提示和人类编写响应的来源,生成十种开源指令微调的大型语言模型响应,并由专家和先进的语言模型进行编辑。Beemo数据集主要用于评估机器生成文本检测器在多作者场景下的表现,旨在解决文本真实性和潜在恶意使用的问题。
The Beemo dataset, developed by Toloka AI, consists of 19.6k text samples covering human-written, machine-generated, and expert-edited content. The dataset includes a diverse range of use cases, such as creative writing, rewriting, summarization, as well as both open-ended and closed-domain question answering. Its construction utilized the No Robots dataset as the source for prompts and human-written responses, generated responses from ten open-source instruction-tuned large language models, and underwent editing by human experts and state-of-the-art language models. The Beemo dataset is primarily designed to evaluate the performance of machine-generated text detectors in multi-author scenarios, aiming to address issues related to text authenticity and potential malicious utilization of generated content.

- 1Beemo: Benchmark of Expert-edited Machine-generated Outputs托洛卡人工智能 · 2024年



