NeelNanda/counterfact-tracing
收藏Hugging Face2022-11-05 更新2024-03-04 收录
下载链接:
https://hf-mirror.com/datasets/NeelNanda/counterfact-tracing
下载链接
链接失效反馈官方服务:
资源简介:
这是一个从ROME论文中的counterfact数据集改编而来的数据集,包含21919个事实关系,每个关系都有两个响应(true和false),旨在用于模型中的记忆编辑和机制解释性研究。数据集的结构包括多个特征(如relation, prompt, target_true等)和分割(如train)。建议使用true和false目标的logit差异来测量,以控制模型识别事实类型的部分。
This is a dataset adapted from the CounterFact dataset in the ROME paper, which contains 21,919 factual relationships. Each relationship is paired with two responses (true and false), and it is designed for research on memory editing and mechanistic interpretability in models. The dataset structure includes multiple features (e.g., "relation", "prompt", "target_true", etc.) and data splits (e.g., "train"). It is recommended to use the logit difference between the true and false targets for measurement, to control for the portion where the model recognizes fact types.
提供机构:
NeelNanda
原始信息汇总
数据集概述
数据集名称
counterfact-tracing
数据集特征
- relation: 字符串类型
- relation_prefix: 字符串类型
- relation_suffix: 字符串类型
- prompt: 字符串类型
- relation_id: 字符串类型
- target_false_id: 字符串类型
- target_true_id: 字符串类型
- target_true: 字符串类型
- target_false: 字符串类型
- subject: 字符串类型
数据集拆分
- train:
- 数据量: 3400668字节
- 示例数量: 21919
数据集大小
- 下载大小: 1109314字节
- 数据集总大小: 3400668字节



