dleemiller/FineCat-NLI

FineCatNLI数据集是一个用于自然语言推理（NLI）的英文单语数据集，包含约100万个样本。数据集通过将七个数据集（包括SNLI、MNLI、WANLI、NLI-FEVER、ANLI、LingNLI等）进行拼接，并使用ModernBERT-large模型进行训练和筛选，以去除容易的样本和错误标签。数据集的筛选系统评估每个前提-假设对在五个独立的质量维度上，以确保数据集的完整性。数据集的标签格式为整数，其中0表示蕴含，1表示中立，2表示矛盾。数据集的每个示例包含前提文本、假设文本、分类标签和原始源数据集标识符。FineCatNLI is an English monolingual dataset for Natural Language Inference (NLI), containing approximately 1 million samples. The dataset is created by concatenating seven datasets (including SNLI, MNLI, WANLI, NLI-FEVER, ANLI, LingNLI, etc.) and training with the ModernBERT-large model to remove easy samples and incorrect labels. The datasets screening system evaluates each premise-hypothesis pair across five independent quality dimensions to ensure dataset integrity. The datasets label format is integer, where 0 represents entailment, 1 represents neutral, and 2 represents contradiction. Each example in the dataset includes premise text, hypothesis text, classification label, and original source dataset identifier.