New finding: The SHAPE of per-layer SAE feature coherence distributions discriminates hallucination better than any scalar summary (K, C, C_delta). Hallucination is not a depth shift — it is a coheren

Zenodo2026-04-09 更新10

Comparison of feature importance across different prediction models (COM, MR, MV, MW, PO).

特征重要性评估

预测模型

Comparison of feature importance across different prediction models (COM, MR, MV, MW, PO).

Figshare2025-12-11 更新10

Feature importance (Importance) and p values for features that achieved a feature importance greater than 0.1% in the Japanese RF.

特征重要性评估

随机森林模型解释

Feature importance (Importance) and p values for features that achieved a feature importance greater than 0.1% in the Japanese RF.

Figshare2023-01-04 更新20

llm-xray-lesion-scans

模型可解释性

模型鲁棒性分析

LLM Lesion X-Ray 数据集是一个已撤回的历史数据集，包含六个约1B参数的语言模型的层损伤扫描结果。该数据集由 Tetracta AI 使用单一预注册协议生成，仅包含测量数据（KL散度值），不包含模型权重。数据集的背景在于：传统基准测试只能评估模型的知识输出，而无法揭示模型内部结构；本数据集通过逐层损伤（剪枝或加噪）并测量输出分布变化，试图回答“模型是如何构建的”这一问题。数据内容包括

Hugging Face2026-09-06 更新10