The engine's few-shot + evaluation examples ("golds") live in state/examples/<scope>/<specialist>.jsonl.
Most are harvested from public sources so each label is ground truth someone else recorded (see
docs/provenance/ for the exact endpoints, counts and scores). Third-party data stays under its own
license — listed here — not under the code's MIT license (LICENSE).
| Scope · files | Source | License / terms | Credit |
|---|---|---|---|
art · classify |
The Metropolitan Museum of Art — Open Access API (catalog fields) | CC0 | optional: The Met Collection API |
baseline · classify, rank, route |
written for this repo (synthetic) | — | — |
baseline · verify |
public GitHub issue titles + engine-drafted replies | short titles; drafts are engine output | — |
acme_corp · respond |
written for this repo (synthetic tenant demo) | — | — |
communities · classify, match |
Stack Exchange network — question titles only | CC BY-SA 4.0 (posts before 2018-05-02: CC BY-SA 3.0) | "Question titles from the Stack Exchange network" |
cooking · classify |
TheMealDB — dish name, ingredients, cuisine | TheMealDB terms of use | "Recipe data from TheMealDB (themealdb.com)" |
economics · classify |
FRED — economic series titles (no data values) | FRED API terms | "This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis." |
general · match |
Wikipedia — redirect / alias name pairs | facts | — |
general · summarize |
Wikipedia — article lead text | CC BY-SA 4.0 | "Text from English Wikipedia; see each article's history for authors" |
government · classify |
Federal Register — document titles + abstracts | U.S. government work, public domain (17 U.S.C. §105) | — |
history · extract, sequence |
Wikipedia — "On this day" event text | CC BY-SA 4.0 | "Text from English Wikipedia; see each article's history for authors" |
legislation · classify, forecast |
Congress.gov — bill titles, CRS policy area, outcome | U.S. government work, public domain | — |
nutrition · classify |
USDA FoodData Central — food descriptions | CC0 | optional: USDA FoodData Central |
science · classify, summarize |
arXiv API — titles, abstracts, categories | arXiv metadata: CC0 | optional: arXiv.org |
sports · forecast |
football-data.co.uk — teams, date, full-time result | match results (facts) | football-data.co.uk |
weather · classify, extract, verify |
U.S. National Weather Service — alert text | U.S. government work, public domain | NOAA / National Weather Service (text unaltered; not an official alert) |
Share-alike. The files drawn from Wikipedia and Stack Exchange (communities/classify.jsonl,
communities/match.jsonl, general/summarize.jsonl, history/extract.jsonl, history/sequence.jsonl)
are adaptations of CC BY-SA content and are released under CC BY-SA 4.0
(https://creativecommons.org/licenses/by-sa/4.0/) — reuse them under that license, with the credit above.
Not in this repo. Some golds the engine was developed with can't be redistributed — verbatim GitHub
issues and maintainer replies, App Store reviews, full Stack Exchange question bodies, and TheMealDB recipe
instructions. They live in a separate private corpus (private/examples/, loaded automatically if
present). A clone of this repo runs on the public golds only, so the evaluation numbers that depended on
private golds are marked as such in docs/provenance/. An earlier film domain sourced from TMDB was
deleted outright: TMDB's API terms prohibit using its content with an AI application.