rockypod/ public-a11y-coder

6 4 days ago

Open-source accessibility coding assistant for the public sector. WCAG 2.2 Level AA conformance, Drupal 11, PHP 8.3, Drush 12, Python 3.12, and Playwright (TypeScript) with both axe-core and Siteimprove Alfa.

4b 14b 
ollama run rockypod/public-a11y-coder:4b

Models

View all →

2 models

public-a11y-coder:4b

2.5GB · 40K context window · Text · 4 days ago

public-a11y-coder:14b

9.0GB · 40K context window · Text · 4 days ago

Readme

Open-source accessibility coding assistant for the public sector.

Trained under privacy-first constraints — no PII, no real agency URLs, no scraped production content. Designed for self-hosted use by government agencies and public-sector developers.

Models

Tag Size Exam (30Q) Best for
rockypod/public-a11y-coder:4b 2.5 GB 73.3% (22.0/30) Portable demos, 8 GB RAM VMs, non-technical explanations
rockypod/public-a11y-coder:14b 8.4 GB 76.7% (23.0/30) Daily-driver technical work, OpenWebUI on agency/homelab hardware

Quick start

# 14B (recommended for technical use)
  ollama pull rockypod/public-a11y-coder:14b
  ollama run rockypod/public-a11y-coder:14b

# 4B (portable demo)
  ollama pull rockypod/public-a11y-coder:4b
  ollama run rockypod/public-a11y-coder:4b

Evaluation

Both variants evaluated on a published 30-question exam covering Drupal 11, PHP 8.3, Drush 12, Python 3.12, Playwright +
  axe-core, Playwright + Siteimprove Alfa, and WCAG 2.2 AA (including the 9 new 2.2 success criteria). Scored Full (1.0) /
  Partial (0.5) / Fail (0.0) per question.

  ┌───────────────────────┬───────┬────────────────────────────┐
  │         Model         │ Score │ Lift vs untrained baseline │
  ├───────────────────────┼───────┼────────────────────────────┤
  │ qwen3:4b (baseline)   │ 45.0% │ —                          │
  ├───────────────────────┼───────┼────────────────────────────┤
  │ qwen3:14b (baseline)  │ 53.3% │ —                          │
  ├───────────────────────┼───────┼────────────────────────────┤
  │ a11y-public-coder:4b  │ 73.3% │ +28.3%                     │
  ├───────────────────────┼───────┼────────────────────────────┤
  │ a11y-public-coder:14b │ 76.7% │ +23.4%                     │
  └───────────────────────┴───────┴────────────────────────────┘

Full rubric, per-question responses, and grades: https://github.com/rockypod/public_a11y_coder

What it knows

- Drupal 11 — render arrays, hooks, Twig 3, custom modules, accessibility patterns
- PHP 8.3 — typed properties, constructor promotion, Drupal service DI
- Drush 12 — CLI usage, custom #[CLI\Command] attribute commands
- Python 3.12 — pathlib utilities, subprocess, pymupdf, BeautifulSoup4 alt audits
- Playwright (TypeScript) — @axe-core/playwright AxeBuilder pattern, @siteimprove/alfa-playwright Audit.run pattern
- WCAG 2.2 AA — all 9 new success criteria (2.4.11, 2.4.12, 2.5.3, 2.5.7, 2.5.8, 3.2.6, 3.3.7, 3.3.8, 3.3.9), cites SC
  numbers in responses

Privacy note

Trained on official public documentation only (drupal.org, w3.org/WAI, playwright.dev, drush.org, php.net, docs.python.org).
No PII, no real agency URLs, no scraped production content. Self-hosted deployment recommended — no prompts leave your
network.