Recipe-Scaling Dataset (from real seeds)

Synthetic recipe-scaling dataset built from license-clean seed recipes (Wikibooks Cookbook, CC-BY-SA) scaled by the deterministic construction engine distilled from The Professional Chef (CIA).

• Seeds: 3782 base recipes (Wikibooks Cookbook, CC-BY-SA).
• Method: each base recipe scaled to several serving counts via the engine (scale path); arithmetic is exact, non-linear corrections applied per role.
• Verification: every record passed deterministic checks (correction bounds, linear/non-linear role semantics, rounding-exactness, valid units).
• Fields: id, method, dish, base_servings, target_servings, scale_ratio, ingredients[{name, role, raw, correction, quantity, note}], reasoning, nl.
• Stats: 22692 generated -> 22692 verified records.
• Files: recipe_scaling.jsonl (full, git-ignored), sample.jsonl (committed).
• License: dataset content derives from CC-BY-SA Wikibooks seeds + engine output; attribute Wikibooks Cookbook contributors. No scraped/pirated sources.

Publish to the Hugging Face Hub

pip install datasets huggingface_hub
huggingface-cli login                      # needs your HF token
python - <<'EOF'
from datasets import load_dataset
ds = load_dataset("json", data_files="data/recipe_scaling.jsonl", split="train")
ds.push_to_hub("<your-username>/recipe-scaling")
EOF