cyu96374/MFT-LLM-binding-axis-multimodel-2026-05
收藏资源简介:
MFT-LLM Binding-Axis Multi-model Backup数据集是一个研究数据集,用于在16种LLM变体中进行道德基础理论的激活导向研究。数据集包含代码、数据和分析结果,主要用于重新运行、重新分析或审计实验。数据集的主要贡献包括ρ校准和绑定轴转移测试。ρ校准使得导向系数可以跨模型比较,而绑定轴转移测试则验证了道德基础理论导向在外部构建(如PVQ价值观和政治意识形态)中的适应性。数据集覆盖了多种模型家族和变体,包括Llama、Qwen、Mistral和Gemma等。数据集的结构包括模型输出、分析结果、可视化数据和日志文件等。
The MFT-LLM Binding-Axis Multi-model Backup dataset is a research dataset designed for activation-steering studies of Moral Foundations Theory across 16 LLM variants. It includes code, data, and analysis artifacts primarily intended for re-running, re-analyzing, or auditing experiments. The datasets main contributions are ρ-calibration and binding-axis transfer tests. ρ-calibration enables cross-model comparison of steering coefficients, while the binding-axis transfer test validates the adaptation of Moral Foundations steering to external constructs like PVQ values and political ideology. The dataset covers multiple model families and variants, including Llama, Qwen, Mistral, and Gemma. Its structure comprises model outputs, analysis results, visualizations, and log files.





