richtext/foodmapper-gte-large

FoodMapper GTE-Large (MLX Format)

This is thenlper/gte-large converted to MLX-Swift safetensors format for use with the FoodMapper macOS application.

Model Description

GTE-Large is a 335M parameter text embedding model that maps sentences to 1024-dimensional dense vectors. It excels at semantic similarity tasks, making it ideal for matching food names across different databases and nomenclatures.

This conversion is optimized for Apple Silicon GPUs via MLX-Swift.

Intended Use

• Semantic food name matching (e.g., matching "granny smith apple" to "Apple, raw, with skin")
• Food database harmonization between USDA FoodData Central, FooDB, and custom datasets
• General text similarity on Apple Silicon Macs

Model Details

Property	Value
Parameters	335M
Embedding Dimension	1024
Max Sequence Length	512
Architecture	BERT
Precision	float16
Format	safetensors

Files

• gte-large.safetensors - Model weights in safetensors format (~670MB)
• config.json - Model architecture configuration
• tokenizer.json - Tokenizer vocabulary and settings
• tokenizer_config.json - Tokenizer configuration
• vocab.txt - WordPiece vocabulary
• special_tokens_map.json - Special token mappings

Usage with FoodMapper

This model is automatically downloaded by the FoodMapper macOS app when first launched. No manual setup required.

Usage with MLX-Swift

import MLX
import MLXNN

// Load weights
let weights = try loadArrays(url: modelURL)
let parameters = ModuleParameters.unflattened(weights)
try model.update(parameters: parameters, verify: .none)

Pooling

GTE models use mean pooling over token embeddings (not CLS token pooling). The attention mask should be applied before averaging:

func meanPooling(_ hiddenState: MLXArray, attentionMask: MLXArray) -> MLXArray {
    let maskExpanded = attentionMask.expandedDimensions(axis: -1)
        .asType(hiddenState.dtype)
    let sumEmbeddings = (hiddenState * maskExpanded).sum(axis: 1)
    let sumMask = MLX.maximum(maskExpanded.sum(axis: 1), MLXArray(1e-9))
    return sumEmbeddings / sumMask
}

Original Model

Based on thenlper/gte-large by Alibaba DAMO Academy.

License

Apache 2.0 (same as original GTE-Large)