ML Interview Practice

Lightweight repo for practicing ML researcher interview tasks (nanoGPT-ish topics plus basics).

Layout

• src/core/: reusable utilities (data, models, training, eval, misc helpers). Includes a modular nanoGPT core for attention/cache/tokenization/norm exercises.
• problems/: self-contained exercises with prompts, starter code, tests, and a notebook.
• tests/: shared fixtures or integration tests.
• scripts/: small helpers (downloads, runners).
• docs/: short guides and topic notes.
• notebooks/: optional notebook copies or scratch space.

Quickstart

• Install deps: pip install -e .[dev] or pip install -r requirements.txt if you prefer.
• Notebook-first flow: start Jupyter, open problems/<name>/notebook.ipynb, edit the starter cell (it writes back to starter.py), then run the bundled pytest cell.
• CLI flow: edit problems/<name>/starter.py, run pytest problems/<name>/tests (or scripts/run_problem.py <name>).

Roadmap

Start with core utilities and the first two problems (conv2d, mnist loop). Then add GPT-ish topics: KV cache, PPO/DPO/GRPO variants, pairwise losses, calibration, context window extensions, tokenization tweaks, normalization choices, optimizer debugging, encoder/decoder/MoE variants.