Machine Learning Practice

This repository is inspired by RezaSi's Golang interview practice

Hands-on coding challenges for Python ML/DL. Each challenge ships with a starter template, a canonical test suite, learning material, and hints. You implement the solution, run the tests locally, and (later) submit it for the leaderboard - a self-hosted, fork-and-PR practice platform in the spirit of go-interview-practice, for machine learning.

Challenges

#	Challenge	Topic	Difficulty
1	Pydantic Models for an ML Training Config	Pydantic v2 validation	Beginner
2	PyTorch Tensor Operations	Tensors: shape, indexing, broadcasting, matmul	Beginner
3	A Small Neural Network in PyTorch	PyTorch nn.Module, training loop	Beginner-Intermediate

More challenges coming.

How a challenge is structured

challenge-N-<slug>/
|-- README.md              # problem statement
|-- metadata.json          # title, difficulty, tags, package requirements
|-- solution_template.py   # skeleton you edit (has TODOs)
|-- test_solution.py       # canonical tests - the source of truth (don't edit)
|-- learning.md            # concept primer with references to official docs
|-- hints.md               # progressive hints
|-- run_tests.sh           # local grader (uv run pytest)
`-- submissions/           # one folder per user
    `-- <username>/solution.py

Solve a challenge

# from the repo root, once: create the environment with uv
uv sync

cd challenge-1-pydantic-ml-config

# set up your submission
mkdir -p submissions/<your-username>
cp solution_template.py submissions/<your-username>/solution.py

# implement submissions/<your-username>/solution.py, then run the tests
./run_tests.sh            # enter your username when prompted

The grader copies your solution.py next to test_solution.py in a temporary directory and runs pytest - your solution passes when all tests pass.

Web UI

A local web app (FastAPI + React/Vite) lets you browse challenges, edit and run code in the browser, and watch the leaderboard.

# 0. From the repo root, install Python deps with uv (creates .venv)
uv sync

# 1. Backend - serves the API (and the built UI) on :20700
uv run uvicorn app.main:app --port 20700 --reload

# 2a. Frontend in dev mode - hot reload on :5173, proxies /api to :20700
cd frontend && npm install && npm run dev      # open http://localhost:5173

# 2b. ...or build once and let the backend serve everything from :20700
cd frontend && npm install && npm run build    # open http://localhost:20700

Regenerate the scoreboards (grades every submission and rewrites each SCOREBOARD.md; the API aggregates those into the leaderboard live):

uv run python scripts/update_scoreboards.py                             # all challenges
uv run python scripts/update_scoreboards.py challenge-3-pytorch-neural-network  # one

Platform layout

app/            FastAPI backend
  challenges.py   discover challenge-* dirs, render markdown
  grader.py       run a submission against tests in a temp dir (pytest)
  scoreboard.py   read SCOREBOARD.md files, aggregate the leaderboard
  main.py         API routes + serves frontend/dist
frontend/       React + Vite + Tailwind + shadcn/ui (neutral theme, light/dark)
scripts/        update_scoreboards.py (batch grader), grade_submission.py (one-user CI grader)

Contribution flow (CI)

Like go-interview-practice, submissions land via fork + PR and the repo automates the rest with GitHub Actions:

• .github/workflows/pr-tests.yml - on every PR to main: checks the PR only touches challenge-*/submissions/<pr-author>/ (files elsewhere need a maintainer to add the manual-approval-granted label), then runs scripts/grade_submission.py for each changed challenge. The final pr-status job is the single check to require in branch protection.
• .github/workflows/auto-merge.yml - every 2 hours, squash-merges any open PR whose pr-status check has been green for 2+ days (tweak WAIT_MS in the workflow to change the wait), then regenerates scoreboards for the challenges it touched.
• .github/workflows/update-scoreboards.yml - on any direct push to main under challenge-*/submissions/** (e.g. a maintainer merging manually), regenerates the affected SCOREBOARD.md files via scripts/update_scoreboards.py and commits them back.

All three commit as github-actions[bot]; no extra secrets are needed beyond the default GITHUB_TOKEN.

Requirements

• Python 3.11+ and uv
• Node 18+ (only for the web UI frontend)

uv sync installs everything (fastapi, a CPU build of torch, pydantic, numpy, pytest, ...) into .venv. All challenges are CPU-only and run in seconds; no GPU required.