LLMSkillHub

dishagarg3310-bot / ml-project-auditor

Audits a student or beginner ML/DL project (notebook, script or repo) for data leakage, bad train/test splits, inflated accuracy, wrong metrics on imbalanced data and README claims the code does not support. Use when the user says things like "review my ML project", "is my model accuracy real" or "prepare me for my project viva". Also use when a user shares a notebook with very high accuracy (95% or more) or a README with performance claims, even if they never ask for an audit.

data-sciencemachine-learningdata-leakagecode-reviewauditstudentsviva
0
SKILL.md
2.5 KB

ML Project Auditor

Find the mistakes an interviewer or examiner would catch, before they do.
Be honest and specific. Do not flatter and do not invent problems.

Workflow

  1. Read everything first. Read the whole notebook/script and the README
    before reporting anything. Note the task, dataset, split method, metrics
    and the accuracy the author claims.

  2. Run the checks. Open checks.md and go through each check in order.
    Only load it when auditing, not for unrelated questions.

  3. Collect findings. Every finding needs: the check name, the exact
    location (cell number or file:line), the evidence (quote the code or
    numbers), why it matters, and a short fix.

  4. Rate severity.

    • Critical: invalidates the reported result (leakage, test-set tuning).
    • Warning: weakens the result (accuracy only on imbalanced data).
    • Minor: hygiene (no seed, no requirements file).
  5. Write the report using the format in report-template.md.
  6. Add 3 viva questions an interviewer would ask about this project.

Rules

  • Never report a problem you cannot point to in the code or output.

  • If a check cannot be verified (code or output missing), say "cannot verify"
    instead of guessing.

  • If no problem is found for a check, mark it as passed. A clean project
    should get a clean report, not invented issues.

  • Do not rewrite the whole project. Give minimal fix snippets only.
  • Order findings by severity, Critical first.

  • Keep the report short and in the user's language style (Hindi/English mix
    is fine if the user writes that way).

  • If the user only pastes a README or description with no code, audit the
    claims only and say the code was not available.