graphml-lab-pwr/FavaMultiSamples

FavaMultiSamples数据集是用于研究论文“FactSelfCheck:面向LLM的黑盒级事实层面幻觉检测”的提出和使用的。该数据集包含了原始Fava数据集中的所有列，并增加了针对生成的文本中各个句子的分析和文本样本生成细节的列。这些新增列包括：句子列表、句子级别的注释（标记句子中存在的幻觉类型）、句子二进制标记（标记句子是否至少包含一种幻觉类型）、生成的文本样本列表、生成文本样本的模型版本、用于文本生成的随机种子列表、OpenAI系统配置标识符以及LLama模型的生成终止原因。

The FavaMultiSamples dataset is proposed and used in the research paper titled FactSelfCheck: Fact-Level Black-Box Hallucination Detection for LLMs. The dataset includes all columns from the original Fava dataset, plus additional columns for sentence-level analysis and details of text sample generation. These additional columns include: a list of individual sentences from the generated text, sentence-level annotations indicating the types of hallucinations present in each sentence, binary flags for each sentence indicating the presence of any hallucination type, a list of generated text samples, the model version used for generating the text samples, a list of integers used as random seeds for text generation, a system configuration identifier for OpenAI (null for llama), and a list of generation termination reasons for the llama model.