Coding Interview Preparation — Data Science Coding Test

A self-extending practice repo for a data science / applied ML internship's online coding test: Data Cleaning & Preprocessing, Data Querying & Retrieval, Exploratory Data Analysis & Visualization, and Machine Learning & Predictive Modeling.

Start with CURRICULUM.md — the detailed checklist of exactly which operations to be fluent in, grounded in public reports of the real assessment format (90–120 min, Python + pandas/numpy/scikit-learn, plus a dedicated SQL section).

How this repo is organized

CURRICULUM.md              topic-by-topic checklist — read this first
quiz/
  quiz.html                 self-contained MCQ practice page — open directly in a browser
  questions.yaml             source data for quiz.html — extend this, then rebuild
  README.md                  how the quiz works, how to add questions
tasks/
  INDEX.md                 auto-generated status table (run tools/build_index.py to refresh)
  001_.../                 one folder per task
    meta.yaml              topic, difficulty, status
    task.md                the brief: business framing, dataset, numbered deliverables
    work.ipynb              <- you work here
    data/
      generate_data.py      regenerates the dataset (synthetic, seeded, reproducible)
      *.csv
    solution_revealed.ipynb <- appears only after you run reveal_solution.py
solutions/
  001_.../solution.ipynb   the hidden worked solution — don't open this until you've tried
tools/
  new_task.py               scaffold a new task folder
  reveal_solution.py         copy a solution into its task folder once you're ready
  run_notebook.py            execute a notebook headlessly (for a quick compare)
  check_task.py               self-grade a submission against a task's expected output (autograded tasks only)
  build_index.py             regenerate tasks/INDEX.md from all the meta.yaml files
  build_quiz.py               regenerate quiz/quiz.html from quiz/questions.yaml
  task_prompt_template.md    the prompt to hand an LLM to write a new task's content
  nbutil.py                  tiny helper for building .ipynb files programmatically

MCQ practice

The real test has a ~9-15 question multiple-choice section (probability, ML fundamentals, model evaluation) separate from the coding problems. Practice that with:

open quiz/quiz.html   # or xdg-open on Linux — no server needed, fully offline

Answering any option reveals the correct answer and explanation immediately. See quiz/README.md for how to add more questions (also LLM-extendable, same pattern as tasks).

Two task styles

• business_framed (tasks 001-004) — a sentence of context, an open-ended deliverable list, graded by your own judgment against the hidden solution.
• exact_io_contract (task 005) — terse, no framing, fixed input/output filenames, lettered sub-steps, strict dtype requirements, and a real autograder. This is the closer match to how the actual platform's questions are phrased (see task 005's brief and notes for a worked example of the classic "NaN forces float64" dtype trap) — prefer this style when generating new tasks unless you specifically want to practice the more open-ended judgment calls.

For exact_io_contract tasks, self-check your output with:

python tools/check_task.py 005

It reports each check (columns, row count, dtypes, values) pass/fail independently, mirroring the platform's per-step partial scoring. Some tasks bundle several small lettered parts into one task instead of many tiny folders (e.g. tasks/008_.../, seven parts a–g) — for those, python tools/check_task.py 008 grades every part in one run, or --part e for just one.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter lab

Workflow for one task

1. Open tasks/00X_.../task.md, read the brief.
2. If data/*.csv isn't there yet, run python data/generate_data.py from inside that task folder (or from anywhere, adjusting the path).
3. Work the problem in tasks/00X_.../work.ipynb, in Jupyter Lab.
4. When you want to check yourself:

   python tools/reveal_solution.py 00X

   This copies the hidden solution into tasks/00X_.../solution_revealed.ipynb, right next to your own notebook. Open both as tabs in Jupyter Lab and compare — or run both headlessly and diff outputs:

   python tools/run_notebook.py tasks/00X_.../work.ipynb
   python tools/run_notebook.py tasks/00X_.../solution_revealed.ipynb

5. Update status: in that task's meta.yaml (in_progress → done), then:

   python tools/build_index.py

   to refresh tasks/INDEX.md.

The solution is never shown until you explicitly reveal it — solutions/ sits outside tasks/ specifically so it doesn't tempt you while browsing the file tree.

Adding more tasks with an LLM

This is meant to grow indefinitely. To add a new task:

python tools/new_task.py --title "Window functions on sales data" \
    --topic data_querying --difficulty medium

This scaffolds an empty tasks/00N_.../ + solutions/00N_.../ pair (folders, meta.yaml, starter task.md, empty work.ipynb/solution.ipynb). It does not write the actual content — hand that off to an LLM (this repo was built with, and is meant to be extended with, Claude Code) using the prompt in tools/task_prompt_template.md, which points it at CURRICULUM.md for topic coverage and an existing task as a style reference.

A reasonable cadence: pick one under-covered subtopic from CURRICULUM.md (cross-check against tasks/INDEX.md's "By topic" counts), generate a task for it, solve it yourself, reveal and compare, repeat.

Design choices worth knowing

• Data is synthetic and seeded, generated by a small numpy/pandas script per task — no external downloads, fully reproducible, and the generator deliberately bakes in the specific data-quality issues or signal the task is testing (not left to chance).
• SQL tasks run against real SQL via DuckDB, directly on the CSVs/ dataframes — no database server to stand up, but genuine SQL syntax, not a pandas approximation of it.
• Solutions favor speed and correctness over polish, deliberately mirroring what the real test rewards (per public candidate reports: ~60–70% correctness was reportedly enough to pass) — don't over-index on making your own code beautiful under time pressure.
• Every solution step traces back to an explicit deliverable. If solution.ipynb does something (e.g. outlier correction, not just the "obvious" cleaning), that step is always its own numbered/lettered item in task.md — never an undocumented surprise. If you ever spot a solution doing something its task.md doesn't ask for, that's a bug — flag it.