code-competitionslisted
Install: claude install-skill TranBaDat2607/claude-code-kaggle-grandmaster
# Code Competitions
In a code competition you submit a notebook; Kaggle re-runs it on the hidden test set (often much
larger than the visible sample) with internet **off**, a hard runtime limit (commonly 9h GPU / 9h
CPU, sometimes much less), fixed hardware, and an output file (`submission.csv` / parquet) or an
evaluation-API loop.
## 1. Day-one checklist
- Make a trivial end-to-end submission (sample submission copy through the full kernel) to learn
the mechanics: where input lands (`/kaggle/input/<slug>/`), the output path, API usage.
- Note the limits in `.kaggle-gm/competition.json` (`code_competition`, `runtime_limit_hours`,
`internet_allowed`, `gpu`).
- Decide the inference budget per model early; it constrains backbone size, folds and TTA.
## 2. Packaging (everything offline)
| Asset | How |
|---|---|
| Model weights | `kaggle datasets create -p kernels/<name>-weights` (dataset-metadata.json), version with `kaggle datasets version -p ... -m "msg"`. Everything in the folder is uploaded; `Dataset version creation error` exits 0, so check the output before pushing a kernel that attaches the new version |
| Your code (`src/`, `kgkit`) | upload as a dataset, `sys.path.append("/kaggle/input/<code-ds>")` |
| Python packages not in the image | `pip download <pkg> -d wheels/ --no-deps` (matching Python/CUDA), upload as dataset, install with `pip install --no-index --find-links /kaggle/input/<wheels-ds> <pkg>` |
| HF models/tokenizers | `save_pretrained` to a folder → d