NateRunsA-LIST/alist-food-text-embedder

A-LIST food-text reader (EmbeddingGemma 2)

Reads a typed food-diary entry ("2 eggs, toast with butter, half an avocado") into 24 nutrition numbers on device, in one forward pass: calories, protein, carbohydrates, fat, fiber, sugar, saturated fat, water (ml), glycemic index, then 15 micronutrients (vitamins A, C, D, E, K, B6, B12, folate, calcium, iron, magnesium, zinc, potassium, selenium, omega-3), in that order and in real units.

Fine-tuned from google/embeddinggemma-2 (text tower only, Apache 2.0) with a regression head, on de-identified food entries labeled by a teacher model. The vocabulary is pruned to the 13,266 pieces seen in training (145M parameters).

Files (v1/)

File	Runtime	Input	Size
FoodTextEmbedder.mlpackage	Core ML (iOS 17+), int8 weights; run on the Neural Engine	input_ids int32 [1,128], right-padded with 0	146 MB
food_text_embedder_int8pc.onnx	ONNX Runtime, int8 per-channel weights	input_ids int64 [1,L], unpadded	169 MB
food_tokenizer.json	Gemma BPE pieces, ids and merges for the pruned vocabulary	—	1.5 MB

Output nutrition float [1,24].

Tokenization: clip to 2,000 code points; text = "task: classification | query: " + entry; U+0020 → ▁; BPE over code points by merge rank (compare pieces by bytes); unknown code points fall back to <0xXX> byte pieces; ids = bos + pieces + eos, truncated to 128 keeping eos.

Precision: EmbeddingGemma 2 activations overflow fp16 on Core ML's CPU path — run the Core ML model on the Neural Engine (or use the ONNX model, which computes in fp32).

Accuracy

Agreement with the teacher within ±25% (or a small absolute slack), 1,500 held-out-user entries:

	Calories	Carbs	Protein	Micros (mean)
Core ML, iPhone 16 Neural Engine (6.4 ms)	0.795	0.776	0.844	0.870
ONNX, Galaxy Z Fold 7 CPU (23 ms)	0.785	0.775	0.847	0.870

Numbers the user typed ("650 cal", "30g protein") should override the model's estimate for that field.