Model Specs Experiment

🧪 Draft — hf-model-spec-0.1 · A working-draft pattern for declarative model specs as Hub artefacts. Not an official HF convention. Feedback welcome.

What this is

A small experiment in model specs — short, declarative descriptions of a desired model (purpose, inputs, outputs, target quality, size budget, data) that an agent (Claude Code, ml-intern, Codex, etc.) can read and execute end-to-end. The spec describes what model should exist; the agent works out the how.

Each spec is a Hugging Face dataset repo with a hf-model-spec-0.1 README. The format is intentionally minimal — YAML frontmatter for machine-readable fields, markdown body for the narrative that doesn't fit in YAML.

Example specs

• model-specs/blbooksgenre-spec — text classification (Victorian book titles → Fiction/Non-fiction)
• model-specs/index-card-blank-detector-spec — image classification (archival index cards → blank/content; retro-fit of a shipped model)
• model-specs/smol-hub-tldr-v2-spec — generative summarisation (HF model/dataset cards → 60-token TL;DR; in-progress)

Feedback

Open a discussion on any spec repo's Community tab. The format is fully expected to change.