Library of Alexandria
A curated cross-cultural mythology, folklore, and religion corpus. Real myths as recorded and believed, in their original languages. Built as a training substrate for local models, not a trivia dataset.
The principle
Most mythology datasets are thin. A few hundred Wikipedia articles, some public domain translations. Fine for trivia, useless for training a model to reason about belief systems. The Library runs on one rule: primary texts, as recorded and believed, with the original languages, not retellings. Sites that opted out of AI training, like sacred-texts.com, were deliberately not crawled. Their robots file says no, so the answer is no.
The cull came first
The honest part of this project is what got deleted. The corpus had grown to 172GB and 9 million files, most of it junk. Nine unfiltered Wikipedia dumps eating 106GB. 434 Gutenberg books with the wrong ebook IDs, so the file said Norse sagas and contained Victorian poetry. 524k OCR page fragments. 31k modern Chinese legal documents that had no business being in a mythology corpus. All of it gone, with manifests kept so every deletion is auditable. Down to 67GB and around 675k files actually worth training on.
Then the verified rebuild
What it feeds
The corpus is the substrate for the world-model research thread: can a small model learn to reason about civilizations from primary sources instead of encyclopedia summaries. It also feeds a local RAG setup, so questions about a tradition come back grounded in the actual texts with citations, not hallucinated summaries.