ndlanier/gutsignal-food-ner-coreml

GutSignal Food NER Model (CoreML)

Named Entity Recognition model for extracting food-related entities from text. Optimized for iOS devices in CoreML format.

Model Description

This model extracts structured entities from natural language food descriptions:

Entity	Description	Examples
FOOD	Food items	chicken, rice, apple
QUANTITY	Amounts	2, half, large
UNIT	Measurements	cups, oz, tbsp
COOKING_METHOD	Preparation	grilled, fried, baked
DESCRIPTOR	Modifiers	fresh, organic, raw

Example

Input: "Chicken breast pan fried with avocado oil and mashed potatoes"

Output:

• FOOD: chicken breast
• COOKING_METHOD: pan fried
• FOOD: avocado oil
• FOOD: mashed potatoes

Model Variants

Variant	Size	Recommended For
FoodNER_float32.mlpackage	~253 MB	Development/testing
FoodNER_float16.mlpackage	~253 MB	Production (iPhone 15+)

Performance Metrics

Entity	F1 Score
FOOD	0.9402
QUANTITY	0.9891
UNIT	0.9720
COOKING_METHOD	0.9950
DESCRIPTOR	0.9547
Overall	0.9533

iOS Integration

Requirements

• iOS 16.0+ (iOS 18.0+ recommended)
• iPhone 15 or newer
• Xcode 15+

Quick Start

1. Download FoodNER_float16.mlpackage and supporting files
2. Add to your Xcode project
3. Use the provided Swift integration code

Swift Usage

import CoreML

class FoodNERService {
    private var model: MLModel?
    private var tokenizer: DistilBERTTokenizer?

    init() throws {
        let config = MLModelConfiguration()
        config.computeUnits = .all  // Use Neural Engine

        guard let modelURL = Bundle.main.url(
            forResource: "FoodNER_float16",
            withExtension: "mlpackage"
        ) else { throw NERError.modelNotFound }

        self.model = try MLModel(contentsOf: modelURL, configuration: config)
        self.tokenizer = try DistilBERTTokenizer(vocabPath: vocabURL)
    }

    func extractEntities(from text: String) throws -> [FoodEntity] {
        // Tokenize input
        let (inputIds, attentionMask, tokens) = tokenizer.encode(text)

        // Run inference
        let output = try model.prediction(from: input)

        // Parse entities from output
        return parseEntities(logits: output.logits, tokens: tokens)
    }
}

See INTEGRATION_GUIDE.md for complete implementation details.

Files Included

• FoodNER_float32.mlpackage - Full precision model
• FoodNER_float16.mlpackage - Recommended for production
• vocab.txt - Tokenizer vocabulary (30,522 tokens)
• tokenizer.json - Full tokenizer configuration
• labels.json - Entity label mapping (11 labels)
• ios_metadata.json - Model metadata for integration
• INTEGRATION_GUIDE.md - Complete Swift integration guide

Training Details

• Base Model: DistilBERT (distilbert-base-uncased)
• Parameters: ~67 million
• Max Sequence Length: 128 tokens
• Training Framework: PyTorch + Transformers
• Conversion: ONNX -> PyTorch -> CoreML (via coremltools 9.0)

Inference Performance

• Latency: 50-150ms on iPhone 15 Pro
• Compute Units: CPU + GPU + Neural Engine (ALL)
• Memory: ~300 MB peak during inference

Limitations

• Best performance on common Western foods
• May struggle with very regional or specialized cuisine terms
• Maximum input length of 128 tokens
• Designed for English text only

Privacy

This model runs entirely on-device with no network calls required. All food parsing happens locally, ensuring user privacy.

License

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0)

You are free to:

• Share — copy and redistribute the material
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit
• NonCommercial — You may not use the material for commercial purposes

For commercial licensing inquiries, contact the author.

Citation

@misc{gutsignal-food-ner,
  author = {GutSignal},
  title = {GutSignal Food NER Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ndlanier/gutsignal-food-ner-coreml}}
}

Related Models

• gutsignal-food-parser-llama-3.2-3b-mlx - Full LLM for complex food parsing
• gutsignal-food-parser-tinyllama-1.1b-mlx - Lightweight LLM option