This model is a fine-tuned version of bert-base-uncased on the food domain data Recipe1M+ dataset. Recipe1M+ contains over 1M records of distinct food names with their ingredients and recipes, more details about the dataset can be found on their project website. We used the whole Recipe1M+ dataset with a total of 1,029,720 records, with using 10% of the dataset as an evaluation dataset. Each of the records contains the food name, followed by its ingredients and recipes.

It achieves the following results on the evaluation set:

Loss: 0.6230

Usage

You can use this model to get embeddings/representations for your food-related dataset that you will use for your downstream tasks.

python

from transformers import pipeline

# Your food-related data
food_data = "Hawaiian Pizza"
# Use pipeline for feature extraction
embedding = pipeline(
        'feature-extraction', model='alexdseo/RecipeBERT', framework='pt'
    )
# Mean pooling
food_rep = embedding(food_data, return_tensors='pt')[0].numpy().mean(axis=0)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 128
eval_batch_size: 128
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
0.7914	1.0	13286	0.7377
0.6945	2.0	26572	0.6569
0.6574	3.0	39858	0.6216

Framework versions

Transformers 4.35.0
Pytorch 2.1.0+cu121
Datasets 2.11.0
Tokenizers 0.14.1