AI transparency
A11y-Equitas-8B is itself an AI model. This page discloses the AI systems used to build it, train it, and maintain this site — so you can evaluate provenance and apply appropriate skepticism.
The deployed model
| Item | Details |
|---|---|
| Base model | Qwen3-8B (Alibaba Cloud / Qwen Team, Apache 2.0) |
| Fine-tuning method | LoRA via Unsloth on Rocky Linux GPU server |
| Training data | Synthetic instruction pairs generated by a local Qwen3-Coder-Next teacher model. No cloud APIs used during data generation — all inference ran on private hardware. |
| Training data topics | WCAG 2.1/2.2, Section 508, DOJ Title II, ADA, state mandates (WA WaTech Policy USER-01 [formerly OCIO 188], OR ADA Title II program, CA AB 434), Playwright + axe-core + Alfa testing, neurodivergent accessibility patterns |
Teacher models (data generation)
| Model | Use | Where it ran |
|---|---|---|
| Qwen3-Coder-Next (Qwen Team) | Primary teacher — generated instruction/answer pairs | Local hardware via Ollama (private, no cloud) |
| Qwen3-14B (Qwen Team) | Secondary teacher and validation | Local hardware via Ollama |
Development tools
| Tool | Use |
|---|---|
| Claude Code (Anthropic — claude-sonnet-4-6) | Agentic coding assistant used for site development, accessible component design, and WCAG audit workflows |
| Google Gemini | Generated the A11y Equitas logo image |
What AI was NOT used for
- Legal or compliance advice (all SC citations are human-reviewed)
- Production decisions about deployed model behavior
- Any processing of user data
Limitations
Synthetic training data produced by a teacher model can reflect that model’s errors. Legal citations in particular should be verified against primary sources. This model is a research and guidance tool, not a substitute for qualified legal counsel.
Last reviewed: 2026-06-02.