← ClaudeAtlas

kaggle-campaignlisted

Use when running a Kaggle (or similar) modelling competition end to end - planning a campaign, deciding what to build next, judging whether to blend or rebuild, choosing final submissions, or when a competition score has plateaued. Encodes the phase order, the frontier gauge, and the build-versus-verify budget that separate a top finish from a merely strong one.
ssubbotin/kaggle-research-skills · ★ 0 · AI & Automation · score 69
Install: claude install-skill ssubbotin/kaggle-research-skills
# Kaggle campaign: how to turn a strong finish into a top finish Derived from campaign postmortems and from the write-ups of winning teams. The pattern the postmortems share: offline scores that transfer to the private board as well as anyone's, and a model class that is still well short of the winner's. The lesson is narrow and repeatable: **such a campaign's verification apparatus is first class and its build throughput is not.** Everything below exists to keep those two in the right proportion. ## The one law > **A competition is decided by the strongest single model you build, and only then by everything > you do to it.** Blending, gating, dose tuning, leaderboard probes, risk gates and selection rules move you within a frame. They cannot move you to a better frame. If your frame is wrong, perfect execution of everything else lands mid-field. ## Phase order (do not reorder) ### Phase 0 — Frame census (first 3 days, before any modelling) Read every public notebook and every discussion thread, and write down **the representation each one uses**, not the model name. What is the input tensor? What are its axes? What is the label space? What is the anchor? Representations, not architectures, are what separates leaderboards. Output: a short table of distinct frames with a headroom guess for each. Then pick one frame and build it yourself end to end. Start from its *frame*, and **reuse every public artefact that serves it**: weights, label tables, auxiliary-model checkpoi