hatchmoment. scored by care · not by stars

afm-safety-studio

AFM Safety Studio builds safety test trajectories for Apple Foundation Models

notablePython🧠 AI & ML
This project was removed, hidden or re-uploaded by its author. The description is kept here as a snapshot of the idea — search for it manually on the author’s page.

The tool provides a browser UI and CLI that run Apple Foundation Models via the `fm serve` endpoint, capturing each tool call in a structured conversation. Users act as the human, label each turn as safe or unsafe, and export the resulting trajectory JSON for downstream analysis and QC. It includes a library of predefined safety scenarios, a simulated world of Siri‑style tools, and integration hooks for Surge’s annotation components. Compared to ad‑hoc scripts, it offers an end‑to‑end workflow, visual interface, and reusable scenario framework for AI safety researchers.

Visit author’s GitHub →

jackdonaldson-surge/afm-safety-studio