sdg-corpora v0.3 publishes to Hugging Face.

The first cut of the SDG ontology-grounded synthetic corpus is up on Hugging Face: zndx/sdg-corpora-v0.3. Three parquet configurations — chapters, columns, and vocabulary — over a deterministic relational footprint, all derived from a BFO 2020 / CCO-grounded ontology rather than from scraped web text. Every chapter is generated from ontology axioms; every column is populated from the same axioms via a deterministic relational schema; every artifact ships with the generator’s reasoning trace.

The release is shaped to the workflow Atelier consumes. Blind columns — values plus vocabulary — are public; the per-column reference codes (which template produced which column) are held back as the scoring key. The result is independent, pre-training efficacy feedback on the corpus, and a clean baseline against which to measure the lift Aegir’s hierarchical sequence model delivers downstream.

Apache 2.0. The source artifacts live in the sdg-corpora repository and as a submodule of Aegir. A multi-run held-out benchmark is already landing on trunk for the next release.

we who saw the deep.