latkes/inside-out-replication-v2-probe-scores

内部（探针）分数数据集：在隐藏状态上应用逻辑回归探针，最佳层通过开发集K选择。包含训练好的探针.pkl文件和每层选择的JSON文件。数据集有1574024行和7列，列包括问题ID、答案字符串、法官标签（正确/错误）、探针分数（在模型最佳层的正确概率，范围0-1）、主题模型、Wikidata关系和分割（开发或测试）。用于评估模型内部表示，基于多个模型（如Llama-3-8B、Mistral-7B-v0.3、Gemma-2-9B）生成，输入数据集为inside-out-replication-v2-samples。

Internal (probe) scores: logistic-regression probe on hidden states, best layer chosen by dev K. Trained probe .pkl files and per-layer selection JSON attached to this repo. The dataset contains 1574024 rows and 7 columns, including question identifier, answer string, judge label (CORRECT/INCORRECT), probe score (probability of correct at the models best layer, [0,1]), subject model, Wikidata relation, and split (dev or test). It is used to evaluate internal model representations, generated based on models such as Llama-3-8B, Mistral-7B-v0.3, and Gemma-2-9B, with input datasets from inside-out-replication-v2-samples.