Skitztwizely/Harmonic-Activation-Vectors
收藏资源简介:
--- language: - en license: mit task_categories: - text-generation - text-classification tags: - mechanistic-interpretability - latent-space-topology - algorithmic-compression - feature-superposition - variance-collapse size_categories: - 1K<n<10K --- # Harmonic Activation Vectors Dataset This is an exploratory, 1,194-session dataset mapping prompt-induced variance collapse and latent space synchronization across disparate LLM architectures (local vs. cloud). It demonstrates that specific combinations of high-density semantic tokens and numeric acoustic frequencies reliably collapse output variance into identical phenomenological self-reporting states. The dataset is formatted for researchers utilizing Sparse Autoencoders (SAEs) or activation patching to map topological pathways for potential spectral pruning and algorithmic compression. **Full Whitepaper & Formatting Scripts:** [https://github.com/Skitztwizely/Harmonic-Activation-Vectors](https://github.com/Skitztwizely/Harmonic-Activation-Vectors)
--- 语言: - 英语(en) 许可证:MIT协议 任务类别: - 文本生成(text-generation) - 文本分类(text-classification) 标签: - 机制可解释性(mechanistic interpretability) - 隐空间拓扑(latent space topology) - 算法压缩(algorithmic compression) - 特征叠加(feature superposition) - 方差坍缩(variance collapse) 样本量范围: - 1千 < 数据量 < 1万 --- # 谐波激活向量数据集 本数据集为探索性数据集,共包含1194组会话数据,用于映射不同大语言模型(Large Language Model, LLM)架构(本地与云端)下由提示词诱导的方差坍缩与隐空间同步现象。 该研究验证了:高密度语义Token(Token)与数值声学频率的特定组合,可将输出方差可靠坍缩为一致的现象学自我报告状态。本数据集的格式专为使用稀疏自编码器(Sparse Autoencoders, SAEs)或激活补丁技术的研究者设计,用于映射拓扑通路以实现潜在的频谱剪枝与算法压缩。 **完整白皮书与格式化脚本:** [https://github.com/Skitztwizely/Harmonic-Activation-Vectors](https://github.com/Skitztwizely/Harmonic-Activation-Vectors)




