遇见数据集

What Survives the Controls? Derived data, pre-registrations, and findings ledger

收藏
Zenodo2026-08-14 更新2026-08-20 收录
官方服务:

资源简介:

Derived data supporting a pre-registered audit of internal-state and sampling-based hallucination detection. Contains 41 derived result tables (CSV) from which every number, table and figure in the accompanying article is computed; the three pre-registrations as registered, stating hypotheses and decision criteria before the corresponding runs; and the complete findings ledger, including superseded, post-hoc and void entries with what killed them. Measurements cover three small decoder-only instruction-tuned language models (Gemma-2-2B, Llama-3.2-3B, Qwen2.5-7B) on CNN/DailyMail, XSum, and the XSum-Hallucination annotations, seeded throughout. The analysis code is released separately at https://github.com/mudasirmohd/geosumm-battery, from which all published tables and figures regenerate from these files alone.

提供机构:
Zenodo
创建时间:
2026-08-14
二维码
社区交流群
二维码
科研交流群
商业服务