Culinary data, precisely calibrated.
Training data for food AI models, with real yields, fat absorption and allergens.
Sample record for one dish. Each number is tied to what happens in the pan.
{ "dish": "Patatas a lo pobre", "potato_ap_g": 1000, "potato_ep_g": 820, "evoo_ml": 150, "absorbed_pct": 30.6, "evidence": "calculated" }
→ 820 g edible portion
30.6% absorbed
Built for teams training food AI models: AI labs, nutrition platforms, smart appliance makers and food R&D.
Recipe text is not a cooking process
A recipe says what goes in the pan. It does not say how much of each ingredient survives prep, or how much oil the food takes up. Models trained on text alone guess at those numbers. We record them.
Yields
Every ingredient has a purchased weight and an edible weight. 1,000 g of potatoes is not 1,000 g on the plate.
Fat and moisture
We account for oil absorbed in frying and water lost in cooking, which text-only recipes leave out.
Allergens and region
Each dish is checked against the 14 EU allergens and tagged with the regional version it comes from.
Three ways to work with us
Start with a small pilot. Scope, acceptance criteria and pricing are agreed per project.
Core data
Recipe extraction, ingredient normalization, standard units, structured schemas and baseline quality checks.
Request a pilotPhysics and compliance
Everything in Core data, plus purchased-to-edible yields, cooking yields, moisture and fat transfer, regional variants and allergen annotation.
Request a pilotGold ground truth
Kitchen measurements where commissioned, image and video review, detailed process mapping, specialist review and full QA reporting.
Request a pilotHow we get to a number, and how sure we are
- Intake and identifyWe record the source and resolve each ingredient to one standard identity.
- Standardize and analyzeUnits are normalized and the cooking process is broken into steps.
- Calculate or measureWe calculate yields and absorption, or weigh them in a kitchen when the project calls for it.
- Review and reportAutomated checks run first. A human expert then resolves any discrepancy. Each record ships with its sources, uncertainty and QA report.
Every value carries an evidence level
Calculated estimates are never presented as laboratory measurements.
Cooks who test software
YieldStandard was founded by two people who have worked in professional kitchens and food businesses, and who have spent years testing enterprise software and evaluating AI data. Every record is made and reviewed by people with kitchen experience. We built it because food data is usually treated as text, and food is physical.
Amy Pickle
Co-founder. Software engineer with fourteen years in the field, and a professional cook and cookbook author.
Hannah Pickle
Co-founder. Leads photo annotation and visual review.
Request a pilot
Tell us what your model needs to learn about food. We will reply with a proposed scope.