The open source AI stack is already strong, but it's fragmented, duplicative, and hard to see as a whole. The Open Source AI Gap Map makes it legible: a living map of what exists across the stack, how open each piece is, how widely it's used, and where the gaps are — so the community can see where to build, where to invest, and where to open things up.
This repo is the data behind that map. Everything is curated YAML in sources/; a
deterministic pipeline validates it, serializes it to build/notebook_data.json, and renders
the published map.
The map is a public, iterative effort, and community curation is the point. There are two ways in:
-
File an issue (no code needed): suggest a product, report an error, or propose a category.
-
Open a pull request editing
sources/. Adding a product is a handful of small YAML files: a product record, a score with citations, and an entry in one category roster and one organization roster. CONTRIBUTING.md has the full recipe and the scoring rubric.
A few ground rules:
- Every score cites a primary source. The map excludes anything it can't verify against one.
- Don't hand-edit generated files.
build/notebook_data.jsonandnotebooks/ai-stack-map.pyare regenerated by a bot on merge; PRs that touch them are blocked. - CI runs on every PR (
build.validate+pytest), so you'll know quickly if something is off.
Each product is graded on three independent, multi-source axes:
- Openness — a 0–5 grade against openness frameworks (the Model Openness Framework for models, OSI license classes for software), not a yes/no. The open-source vs. open-weights distinction is the one the map exists to draw.
- Adoption — real usage (downloads, active users, deployments), not GitHub stars.
- Capability — community benchmarks where they exist, feature coverage where they don't.
Adoption and capability blend into a single overall score per product, banded as Leading (score ≥ 4.5) or Strong (4.0 ≤ score < 4.5). Categories then roll up from their products into a Maturity Stage (0 Void → 5 Mature) plus a set of gaps naming what the open ecosystem still lacks. The taxonomy and openness framework descend from the 2024 Columbia Convening on Openness in AI.
The complete method — the stage and gap formulas, per-axis sources, and stated limitations — is in
docs/methodology.md, the hand-authored source of truth (also rendered
on the site).
One YAML file per record: four concerns plus the single sources/taxonomy.yaml manifest. This is
what you edit.
| Path | Contains | Key rule |
|---|---|---|
sources/organizations/ |
Org metadata and a products: roster |
Each product slug appears in exactly one org roster |
sources/categories/ |
Category definition (weights, strapline, …) and an ordered products: roster |
Order = display order; each product in exactly one category |
sources/products/ |
Product record (name, type, description, typed artifact URLs) |
Org membership lives in the org file, not here |
sources/scores/ |
Per-product openness, adoption, capability |
Every non-null score value needs a sources: citation |
sources/taxonomy.yaml |
Arc grouping + display order; the three arcs are the Columbia ontology layers | Every category appears in exactly one arc |
sources/allowlists/ |
undigested_sources.txt, the sources grandfathered before the digest discipline |
Anything not listed must carry a content_sha256; the list only shrinks |
sources/snapshots/ |
long_tail.json, the frozen dedup counts |
Frozen universe counts; roster counts are derived at build time |
Category slugs use underscore form (base_pretrained); product and org slugs use hyphenated
kebab-case (llama-3-1, allen-ai). Artifact keys on products (only those that apply): github,
npm, pypi, crates, go, huggingface_model, huggingface_dataset. JSON Schemas for every
file type live in docs/schemas/.
Requires Python 3.12+ and uv. No API key is needed to edit sources or validate.
uv sync
uv run python -m build.validate # schema + cross-file checks; must print "0 error(s)"
uv run pytest -q # optional; the same suite CI runsTo preview the generated map locally (the output is for preview only — don't commit it):
uv run python -m build.serialize # sources/ → build/notebook_data.json
uv run python -m build.render # → notebooks/ai-stack-map.py
uv run marimo export html notebooks/ai-stack-map.py -o /tmp/preview.htmlWarehouse queries (via pyoso) need OSO_API_KEY; with direnv, place it in .env and it loads automatically.
| Path | Role |
|---|---|
sources/ |
Curated YAML you edit: organizations, categories, products, scores, plus taxonomy.yaml, allowlists/ and snapshots/ |
build/ |
Deterministic validate → serialize → render pipeline |
notebooks/ |
Generated ai-stack-map.py + standalone companion notebooks |
docs/ |
JSON Schemas, contributor guides, methodology, and maintainer runbooks |
warehouse/ |
UDM SQL and ingest fetchers for adoption / activity signals |
skills/ |
Agent skills mirroring the contribution recipes |
tests/ |
pytest suite for build helpers |
Guides worth knowing: openness scoring,
adoption scoring, gap analysis,
and query conventions.
See AGENTS.md for agent-oriented project context, the full directory map, and
maintainer operations. How the data assets fit together — namespaces, gates, and the repository
boundary — is in docs/architecture/.
Code and data are MIT licensed.