Design Original ML Benchmark Challenges (ML Experts ONLY)

I run an agent-evaluation platform and need a steady stream of fresh, production-grade machine-learning challenges that can be dropped straight into our test harness. For each accepted task I pay USD 200 once the repository clears a brief code and licensing review. Scope of each challenge • Build or source a commercially usable dataset (public-domain, permissive-license, or self-generated). • Write deterministic prepare.py and grade.py scripts compatible with Python 3.10; they must run headless on Linux via a single command. • Document the problem in a concise README that explains objectives, data schema, evaluation metric, and submission format. • Provide at least five clear, task-specific rubric checks the grading script can expose as JSON. • Supply a clean, well-commented reference solution in a Jupyter notebook that reaches a reasonable benchmark score. • Keep everything reproducible and free of copyrighted Kaggle material or LLM-generated filler. Technology expectations Pandas, scikit-learn, XGBoost or LightGBM are welcome for classic tabular work; PyTorch or TensorFlow may be better for vision or NLP—use whatever best serves the task while keeping dependencies lightweight. Acceptance criteria 1. Repository installs with pip and runs end-to-end on a GitHub-hosted CI test. 2. Grade script yields identical scores across repeated runs. 3. Dataset license file included and compatible with commercial redistribution. 4. Reference solution executes without manual tweaks and passes all rubric checks. If you enjoy crafting Kaggle-style problems and have a track record of well-engineered notebooks or GitHub projects, send over a link to past work and let’s line up the first challenge.