pyvene/axbench-concept10

Concept10数据集是为了监督字典学习（SDL）而设计的，包含了10个随机选取的概念的训练和推理数据。这些数据来自于GemmaScope概念列表的Gemma-2-2B-it和Gemma-2-9B-it版本的不同层级。数据集分为文本、代码和数学三种类型，每种类型都有输入指令和相应的模型或LLM生成的输出。输出中会包含一个概念，如果没有概念则标记为EEEEE。此外，数据集还区分了正负样本，并为指令调整模型设置了特定类别。每个子集包含216个负面例子和720个正面例子。

The Concept10 dataset is designed for Supervised Dictionary Learning (SDL) and contains training and inference data for 10 randomly selected concepts from the GemmaScope concept list at different layers of Gemma-2-2B-it and Gemma-2-9B-it. The dataset is divided into three genres: text, code, and math, each with corresponding input instructions and outputs generated by the model or LLM. The output includes a concept, marked as EEEEE if there is no concept. The dataset also categorizes positive and negative samples and is specifically tailored for instruction-tuned models. Each subset includes 216 negative examples and 720 positive examples.