robertteleng/ingredient-detector-yolo-coreml

Ingredient Detector YOLO · CoreML

YOLO26n fine-tuned to detect 53 classes of fridge and kitchen ingredients, exported to CoreML (.mlpackage) to run on the Apple Neural Engine inside an iOS app.

Unlike the other conversions in this account, this one is a trained model, not just a format change.

Results

Trained for 100 epochs on an RTX 5060 Ti. The target was set before training: mAP@0.5 above 0.60.

Metric	Value
mAP@0.5	0.665
mAP@0.5:0.95	0.497
Epochs	100
Classes	53

Training data

Roboflow "Fridge Object" v3 — 9,874 images, 53 classes. No class filtering was needed. Check that dataset's own terms before redistributing the data.

Reproducing it

With robertteleng/coreml-forge:

yolo detect train model=yolo26n.pt data=data/ingredients.yaml epochs=100
uv run python scripts/export_ingredient_yolo.py --weights runs/.../best.pt

Limitations

• Not evaluated on real fridge photos, only on the dataset's own validation split. Expect worse results on cluttered shelves, poor lighting or partially hidden items.
• mAP@0.5 of 0.665 means roughly a third of detections are wrong at that threshold. Fine for suggesting ingredients to a person who can correct them; not fine for anything automatic.
• Per-class performance was not broken down, so rare classes may be far weaker than the average.
• FP16 export: numerically slightly different from the PyTorch weights.

Licence — read this before using it

AGPL-3.0, inherited from Ultralytics YOLO26. This is not a formality: a model fine-tuned from Ultralytics weights carries AGPL-3.0, so shipping it inside a closed-source application requires a commercial licence from Ultralytics. If your use is commercial and your source is not open, talk to them first.