pimsoje/recipe-ner-small-bert-bio

Recipe NER Small BERT

This repository contains a compact BERT-style token classification model trained from scratch on a BIO-annotated recipe corpus.

What the model predicts

• —NAME — ingredient names
• —QTY — quantities
• —UNIT — measurement units
• —COMMENT — modifiers such as chopped / fresh / canned
• —INDEX and RANGE_END — rare auxiliary labels from the source markup

Test metrics

• —F1: 0.8828
• —Precision: 0.8762
• —Recall: 0.8895
• —Accuracy: 0.8960

Usage

from transformers import pipeline

pipe = pipeline(
    "token-classification",
    model="recipe_ner_small_bert",
    tokenizer="recipe_ner_small_bert",
    aggregation_strategy="simple",
)

pipe("2 cups flour\n1 cup sugar\n1 tbsp olive oil")