jkminder/model-raising-interp-prepared
收藏资源简介:
这是一个多标签数据集,用于训练线性探针,以探测模型在阅读文档时如何表示35节Value Charter。数据集包含100,000行数据,每行包括模型阅读的文档片段(text)和对应的charter章节标签(label_ids),标签以JSON字符串格式存储,例如[2.1, 3.3]。数据集经过清理、去重和温和重平衡处理,以提高稀有章节在训练集中的代表性。数据集的目的是比较不同模型变体(如经过charter预训练与普通预训练的模型)在解码charter信息时的能力。数据集还包括章节分布表格,显示每个章节的文档数量,例如章节2.1(Physical Safety)有26,880个文档,而章节4.6(Consent)只有1,189个文档。数据集用于训练分割,评估使用独立的数据集。
This is a multi-label dataset designed for training linear probes to investigate how models represent the 35-section Value Charter during document reading. The dataset comprises 100,000 rows, with each row containing a document snippet read by the model and its corresponding charter section labels (label_ids), which are stored in JSON string format, e.g., [2.1, 3.3]. The dataset has been cleaned, deduplicated, and gently rebalanced to enhance the representation of rare sections in the training set. The primary goal of this dataset is to compare the capabilities of different model variants (such as charter-pre-trained models versus standard pre-trained models) when decoding charter-related information. Additionally, the dataset includes a section distribution table that details the number of documents associated with each section: for example, section 2.1 (Physical Safety) contains 26,880 documents, while section 4.6 (Consent) has only 1,189 documents. The training partition of this dataset is used for model training, while an independent dataset is employed for evaluation.




