Nvidia Kaggle Skill

NVIDIA Kaggle Skill

Purpose

Use this skill for Kaggle competition work: context gathering, writeups, discussions, kernels, local reproduction, submission, and dataset upload.

Do not use it for unrelated ML training, generic notebook editing, general data analysis, or non-Kaggle dataset management unless the user explicitly ties the task to Kaggle.

Inputs

Prerequisites

• Run commands from this skill directory unless a referenced workflow says otherwise.
• Install only the runtime packages needed for the requested workflow.
• Set KAGGLE_API_TOKEN before API, CLI, kernel, discussion, dataset, or submission workflows.
• Confirm local disk space before downloading competition data, kernel inputs, or extracted archives.
• Require explicit user confirmation before sensitive, externally visible actions: competition submissions (each can consume a daily slot), dataset uploads, and creating a public dataset. Treat KAGGLE_API_TOKEN as a secret — never print, log, or echo it.

Runtime Dependencies

Install only the packages needed for the requested task into the current environment, then run scripts with python.

Kaggle API, CLI, kernels, discussions, datasets, competition pages, and writeups:

if command -v uv >/dev/null 2>&1; then
  uv pip install httpx kaggle kagglesdk nbformat pydantic python-dotenv rich
else
  python -m pip install httpx kaggle kagglesdk nbformat pydantic python-dotenv rich
fi

For API/CLI tasks, verify credentials before calling Kaggle:

: "${KAGGLE_API_TOKEN:?ERROR: KAGGLE_API_TOKEN environment variable is not set}"

Workflows

Use this workflow catalog to choose the right path. Run the direct script commands for quick tasks. For workflows that point to another markdown file, read that file only when the request needs that workflow.

Prefer the runtime's run_script helper when it exists, for example run_script("scripts/fetch_competition_info.py", args=["titanic"]). Otherwise run the equivalent python ./scripts/<script>.py ... command from this skill directory.

Competition Details

Use this when the user asks to retrieve or summarize a Kaggle competition overview, rules, evaluation, timeline, or dataset description.

Fetch overview:

python ./scripts/fetch_competition_info.py <competition-slug-or-url>

Fetch dataset description:

python ./scripts/fetch_dataset_info.py <competition-slug-or-url>

The scripts accept a bare competition slug or https://www.kaggle.com/competitions/<slug> URL and extract the slug automatically. Convert output to markdown when the user asks for saved documentation, using {slug}_competition_overview.md and {slug}_dataset_description.md in the current working directory.

Research Brief

Use this when the user asks you to research a competition and write a strategy brief in natural terms (e.g. "research this competition and brief me, with links and a few charts"). You chain the skill's individual research workflows yourself, write your own analysis/plotting code, and produce the brief. Read ./research-brief.md for the principles that keep the brief accurate, informative, and useful to a reader — how to cite real sources as links, and how to make plots honest and legible (every plotted number traces to what you gathered). These principles live in the skill so the user does not have to spell them out.

Writeups

Use this when the user asks to fetch one writeup, fetch top-k writeups, discover leaderboard writeup links, or summarize solution posts. Read ./writeups.md.

Discussions

Use this when the user asks for Kaggle competition discussions, community insights, questions, tips, or a specific discussion thread.

python ./scripts/discussion_ingest.py <competition_id> [--max-pages N] [--sort-by hotness|votes|comments|created|updated] [--page-size N] [--nofetch-comments]
python ./scripts/discussion_query.py <competition_id> [--search TERM] [--min-votes N] [--author NAME] [--limit N] [--as-json]
python ./scripts/discussion_read.py <discussion_id> [--competition-id ID]
python ./scripts/discussion_db_info.py [competition_id]

Storage:

Always run ingest before query/read if the database is empty. Keep retries bounded if Kaggle rate limits or API shapes change.

Kernels

Use this when the user asks to ingest, query, or read kernels; research top public kernels; fetch kernel scores; or analyze kernel lineage. Read ./kernels.md.

Kernel Setup

Use this when the user asks to download and reproduce a Kaggle notebook locally with its inputs. Read ./kernel-setup.md.