菜单
Scaffold a new tabular-multiclass competition in this kit — download/place data, detect id/target/classes/metric, fill src/config.yaml, pick the fold scheme, and freeze the shared CV split. Use once at the very start, before EDA.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Combine all model OOF predictions into the best validated blend/stack, apply the frozen recipe to test predictions, and write submission.csv. Use after the model agents finish, or any time the OOF pool changes.
Create or validate the shared cross-validation split (artifacts/folds.parquet) that every model must use, and check that it tracks the leaderboard. Use before any model trains, or whenever CV and LB diverge.
Validate a submission.csv against sample_submission before spending a scarce daily submission slot — column names/order, id coverage, prob ranges/sums, NaNs, and canonical class order. Use right before every leaderboard submission.
基于 SOC 职业分类
| name | kaggle-new-competition |
| description | Scaffold a new tabular-multiclass competition in this kit — download/place data, detect id/target/classes/metric, fill src/config.yaml, pick the fold scheme, and freeze the shared CV split. Use once at the very start, before EDA. |
Run this once when starting a fresh competition. It gets the contract in place so every downstream agent composes.
Get the data into data/. Either place train, test, sample_submission there
manually, or (if the Kaggle CLI is configured):
kaggle competitions download -c <competition-slug> -p data/ && (cd data && unzip -o '*.zip')
Inspect the shapes and the target. Load train/test/sample_submission, print:
dtypes, the id column (matches sample_submission), the target column, the unique
classes and their counts, and the prob-column layout of sample_submission (this is
the canonical class order your submission must match).
Identify the metric. Read the competition's evaluation page. Map it to one of:
logloss | accuracy | macro_f1 | qwk (extend src/metrics.py if it's something else).
Getting this wrong wastes the whole effort — confirm it literally.
Pick the fold scheme. Independent rows → stratified. Multiple rows per entity (a
group_col exists and must not split across train/val) → stratified_group. A time column
with test in the future → time. Decide from the data description, not a guess.
Fill src/config.yaml — competition, id_col, target_col, metric, n_folds, fold_scheme, group_col, seed, and the data_dir.
Freeze the split and baseline:
python src/make_folds.py # writes artifacts/folds.parquet — the shared CV
Confirm the printed per-fold class balance looks right. Note the class-prior baseline score
in EXPERIMENTS_LOG.md as the floor to beat.
Hand off to EDA (kaggle-eda agent / prompt 1, Stage 1).
folds.parquet exists — the split is the contract.sample_submission column order = your canonical class order; a silent permutation
destroys logloss even with a perfect model.