Apocalypse Bench

Summary

TypeScript CLI named apocbench for no-tools survival questions. 500-question JSONL bank, SQLite run store, HTML and Markdown reports, optional Rust wiki-search over local Wikipedia. Candidate answers from own knowledge; separate judge model on a 10-point rubric. Research config: 500 questions times 9 model families times 2 conditions. MIT.

Keywords

apocalypse-bench · survival · offline-rag · wikipedia · llm-eval

Topics

LLM evaluation, practical survival knowledge, offline RAG

Research notes

• Primary: GitHub README (MIT). 500-question eval harness, not a substantial training dataset. Discord posted the repo.