FredQuant/behavioral-frequency-signatures
收藏资源简介:
该数据集提供了一个用于大型语言模型(LLMs)的紧凑频谱身份框架——行为频率签名。它通过校准采样、信号构建、频谱分解与黄金比例抽取等方法,从LLMs中提取无需访问模型权重的行为签名,签名大小仅为几百字节,与原始模型大小的比例超过138,000,000:1。数据集内容包括技术白皮书摘要、方法概述、10个不同大小模型(3GB至66GB)的签名大小对比表、提取流程和应用场景(如分布式模型路由、边缘咨询协调等),但未包含实际数据文件或样本。
This dataset provides a compact spectral identity framework for Large Language Models (LLMs) — the Behavior Frequency Signature (BFS). It extracts behavior signatures from LLMs without requiring access to model weights, via methods including calibrated sampling, signal construction, spectral decomposition and golden ratio extraction. Each signature is merely several hundred bytes in size, with a size ratio exceeding 138,000,000:1 relative to the total size of the original model. The dataset covers abstracts of technical whitepapers, an overview of the proposed methodology, a comparative table of signature sizes for 10 models ranging from 3GB to 66GB, extraction workflows and application scenarios such as distributed model routing and edge consultation coordination, but does not include any actual data files or samples.



