Skip to content
Design Bench · Holding AGI accountable

Constitutional training corpus

A dataset pipeline turning a design archive into training, validation, and retrieval data planes for constitutional reasoning. Explicitly a draft — the constitution it's built from is itself unratified pending review.

Design boardinternal designation — Thirstys Standard V4
Design board

Named and scoped. Nothing built yet, or too early to show.

In more detail

313 supervised fine-tuning records, 46 validation records, 521 runtime-retrieval records, and a held-out evaluation set explicitly marked DO_NOT_TRAIN. Its own README states plainly that inclusion in the corpus 'does not ratify, verify, or promote' the V4 Constitution or the V3Q manifest it draws from — draft status is stated outright, not left to be inferred.

Concrete specifics

  • ·313 SFT + 46 validation + 521 retrieval records, held-out eval fixtures marked DO_NOT_TRAIN
  • ·README states outright that corpus inclusion does not ratify the underlying constitution

Directly related

Go deeper