juusopi/grocery-fi-ner

Finnish Grocery NER

spaCy named-entity recognition for Finnish grocery text — quantities, units, products, and notes.

Backed by TurkuNLP/bert-base-finnish-cased-v1; the pipeline is transformer → ner.

This repository holds the ner model only. NER and text classification are trained and published separately, so the matching textcat model lives in its own repository and is versioned independently of this one.

Labels

Entity labels: NOTE, PRODUCT, QUANTITY, UNIT

Performance (dev set)

Metric	Score
F1	0.993
Precision	0.991
Recall	0.995

Per-label:

Label	Precision	Recall	F1
PRODUCT	0.989	0.989	0.989
QUANTITY	0.995	1.000	0.998
NOTE	0.975	0.997	0.986
UNIT	0.999	1.000	1.000

Note that the development set is generated from the same vocabulary as the training set, so these figures measure fit rather than generalisation. Held-out performance is measured against a hand-labelled probe set in the training repository.

Usage

import spacy

nlp = spacy.load("juusopi/grocery-fi-ner")
doc = nlp("500 g omenaa")

for ent in doc.ents:
    print(ent.text, ent.label_)
# 500 QUANTITY
# g UNIT
# omenaa PRODUCT

Model details

• Base model: TurkuNLP/bert-base-finnish-cased-v1
• spaCy version: >=3.8.11,<3.9.0
• Version: 1.2.0

Versions are semantic and describe the output contract: a major bump means the label set changed and consumers mapping those labels must be updated.