Transluce/act_patch_qwen3_8b_counterfact

该数据集包含多个特征，包括层序列、输入令牌、原始延续、消融延续、差异标志、补丁位置（包含反事实位置、反事实文本令牌、干预向量、原始位置和原始文本令牌）、反事实文本、原始目标真实值、反事实目标真实值、可哈希层序列、令牌类型等。数据集分为训练集和测试集，训练集包含120,644个样本，测试集包含14,357个样本。

The dataset includes multiple features such as layer sequence, input tokens, original continuation, ablated continuation, difference flag, patch position (containing counterfactual position, counterfactual text token, intervention vector, original position, and original text token), counterfactual text, ground truth original target, ground truth counterfactual target, hashable layer sequence, token type, etc. The dataset is divided into train and test sets, with the train set containing 120,644 examples and the test set containing 14,357 examples.