# Kaggle CPU runbook 1. Create a Kaggle Notebook with **CPU** accelerator (4 vCPU, 30 GB RAM). 2. Add a Kaggle Secret named `HF_TOKEN` with a token that can write to `xLeverd/chess`. 3. Upload/import `kaggle_distill.ipynb` and run all cells. 4. The notebook downloads Stockfish 19 for Linux, the generator from the dataset repo, builds shards in `/kaggle/working/data`, validates them, and uploads them to `xLeverd/chess`. ## Manual equivalent ```bash pip install python-chess pyarrow huggingface_hub tqdm wget -q https://github.com/official-stockfish/Stockfish/releases/download/sf_19/stockfish-linux-x86-64-universal.tar.gz -O sf.tar.gz mkdir -p sf && tar -xzf sf.tar.gz -C sf SF=$(find sf -type f -name 'stockfish*' | head -1) python distill.py --target 300000 --workers 4 --depth 12 --out data --stockfish "$SF" python upload_hf.py ``` `distill.py` is resumable: it counts existing rows in `data/*.parquet` and generates only the remainder. ## Budget At depth 12 on 4 vCPU, expect roughly 80–150 positions/s aggregate. 100k rows is ~10–20 min, 500k is ~1–2 h. Start with 100k, validate, then scale.