Transluce/act_patch_llama_3.1_8b_counterfact

该数据集包含多个特征，主要用于文本处理和干预分析。特征包括层信息（layer）、输入标记（input_tokens）、原始延续（original_continuation）、消融延续（ablated_continuation）、差异标志（is_different）、补丁位置（patch_position，包含反事实位置、反事实文本标记、干预向量、原始位置和原始文本标记等子特征）、反事实文本（counterfactual_text）、原始目标（gt_original_target）、反事实目标（gt_counterfactual_target）、可哈希层（layer_hashable）和标记类型（token_type）。数据集分为训练集（train）、测试集（test）和完整训练集（train_full），分别包含13,454、5,600和107,130个样本。

This dataset includes multiple features primarily used for text processing and intervention analysis. Features include layer information (layer), input tokens (input_tokens), original continuation (original_continuation), ablated continuation (ablated_continuation), difference flag (is_different), patch position (patch_position, containing sub-features such as counterfactual position, counterfactual text token, intervention vector, original position, and original text token), counterfactual text (counterfactual_text), original target (gt_original_target), counterfactual target (gt_counterfactual_target), hashable layer (layer_hashable), and token type (token_type). The dataset is divided into training set (train), test set (test), and full training set (train_full), containing 13,454, 5,600, and 107,130 samples respectively.