pyre-code-ml-practice — Claude Code skill: ML drills in the browser

What pyre-code-ml-practice does

Pyre Code is an ML practice platform you run on your own machine. The problem set spans 68 entries, from ReLU and softmax up to flow matching, DPO, GRPO and Mamba SSM. You write code in a browser editor and get a pass or fail back right away. Everything is graded on CPU, so no graphics card is involved.

The skill covers three install routes. The quick one runs ./setup.sh, which builds a .venv, pulls in the Python dependencies and prints the command to start things. There is also a conda path pinned to Python 3.11, and a Docker path via docker compose up --build. Once running, the grading API answers on http://localhost:8000 and the web app on http://localhost:3000. Under Docker your progress lives in a volume; docker compose down -v wipes it.

Adding problems is the other half. Each one is a Python module under torch_judge/problems/ holding a description, starter code, a reference solution and the test function. The skill lays out that structure, so you can copy the shape for your own exercise. The built-in AI hint is optional and switches on through AI_HELP_BASE_URL, AI_HELP_API_KEY and AI_HELP_MODEL in web/.env; any OpenAI-compatible endpoint works. With no server-side key, a user can still paste their own in the interface.