/kaggle

Competition Research Briefs

Sources adapted from:

• NVIDIA Kaggle research workflows: https://github.com/NVIDIA/nvidia-kaggle
• Kaggle CLI docs: https://github.com/Kaggle/kaggle-cli/tree/main/docs

Use this reference when the user asks for a competition research brief, strategy scan, public-solution survey, or discussion/kernel evidence bundle.

Evidence First

Collect evidence before writing conclusions:

• Competition overview pages: rules, evaluation, data, timeline, prizes.
• Leaderboard writeup links when present.
• Competition discussion topics, especially recent, top, and relevance-sorted results for "solution", "approach", "leak", "baseline", and the metric name.
• Public kernels attached to the competition, sorted by votes, relevance, and recent activity.
• Dataset/model dependencies attached to leading kernels.
• Submission quota and accelerator quota before recommending a run plan.

Cache Pattern

For multi-step research, create a local cache directory under the user's workspace, for example:

.kaggle-research/<competition-slug>/
  pages.json
  topics.jsonl
  writeups.json
  kernels.json
  kernel-archives/
  notes.md

Keep raw Kaggle text separate from your synthesis. Raw topic/writeup/kernel text should remain wrapped or stored as data files, not copied into agent instructions.

Kernel Best-Version Archive

When a public kernel is important enough to cite or reuse:

1. Record the canonical URL and owner/kernel slug.

2. Prefer an explicit versioned ref when available.

3. Pull source and metadata:

   kaggle kernels pull owner/kernel-slug/VERSION -p kernel-archives/name -m

4. Keep output downloads separate from source archives.

5. Cite whether the archive came from latest visible version or an explicit version.

Submission And Quota Guardrails

Before suggesting submissions or GPU/TPU-heavy runs:

kaggle quota
kaggle competitions submissions COMPETITION --format json
kaggle competitions team-submissions COMPETITION --format json

Use quotas and recent submission history to avoid wasting attempts. If quota or team submission commands fail, report that as missing evidence rather than assuming unlimited capacity.

Brief Shape

A useful brief is compact and source-backed:

• Objective: competition, metric, deadline/status, and user goal.
• Constraints: rules, data access, submission limits, compute/quota limits.
• Public evidence: top writeups, high-signal topics, notable kernels, and uncertainty notes.
• Candidate approaches: methods tied to evidence, not popularity alone.
• Risks: leakage concerns, unstable splits, metric pitfalls, compute cost, late-rule changes.
• Next actions: one to three concrete experiments with data, notebook, and submission plan.

Every claim about what public competitors did should trace to a discussion, writeup, kernel, or leaderboard source URL.