Instruction corpus compiled from TypeSafe's public Jev documentation: 119 rows split across the Choice, Noul and Score primitives, with the captured source files, a standard-library Python builder and a published HuggingFace dataset.
User repository on GitHub
INSTRUCT_JEV - TypeSafe AI Jev / System One instruction corpus (choice/noul/score), compiled by DeckerGUI. 119 rows. Mirrored on HuggingFace.
Repo stats from the GitHub API, cached Oct 2026.Homepage
INSTRUCTJEV — a TypeSafe AI Jev / System One instruction corpus, compiled by DeckerGUI from the public TypeSafe AI documentation. 119 instruction rows · choice 47 / noul 51 / score 21. 128 source files captured · 114 unique documents.
Quoted from the project's README. Read the full README.
Watch out
The corpus is compiled from TypeSafe's public documentation by a third party rather than written by TypeSafe, so it inherits whatever that documentation says about the primitives and can drift from the current docs. Non-text capture folders such as PDF and webpage directories are excluded from version control, so the corpus is Markdown, text and JSON only and anything that lived in a rendered page is missing. There is no evaluation harness here: nothing in the repository scores a model on these rows, and the content is documentation text rather than measured behaviour. The builder rewrites tracked output directories, so expect a large diff after a rebuild. Copy reviewed against the repository README on 2026-10-04, not from running it.
PyModel/jev-judge-mcpPython
ajensenwaud/hermes-jev-pluginPython
geilt/typesafe-cliPython
sumleo/prompt2jevPython
lucianfialho/jev-model-routerPythonPosts from builders shipping with Jev right now.
