README.md

metadata

language: ru
library_name: spacy
pipeline_tag: token-classification
tags:
  - spacy
  - ner
  - token-classification
  - product-classifier
  - russian
license: other
metrics:
  - precision
  - recall
  - f1

Product Classifier NER v2

spaCy NER pipeline for extracting structured product attributes from Russian product titles and catalog nomenclature.

Model Details

• Language: ru
• Framework: spaCy
• Pipeline: tok2vec, ner
• Model path in source project: models/ner-v2/model-best
• spaCy compatibility: >=3.8.14,<3.9.0

Entity Labels

The model predicts the following entity labels:

Label	Description
BRAND	Brand, manufacturer, or vendor-like name.
COLOR	Product color.
DENSITY	Density value or density class.
DIAMETER	Diameter value.
LENGTH	Length value.
MODEL	Model, article-like model name, or product series.
POWER	Power rating.
PRODUCT_TYPE	Main product type or product noun phrase.
PROTECTION_RATING	Protection rating such as IP class.
QUANTITY	Count, pack quantity, or quantity-like value.
SIZE	Size, dimension, or size class.
STANDARD	Standard, specification, or normative class.
VOLUME	Volume value.
WEIGHT	Weight value.

Evaluation

Metrics are copied from the trained spaCy meta.json:

Metric	Value
Entity F1	0.8055687391
Entity precision	0.8131587352
Entity recall	0.7981191223

Usage

Install the model package from Hugging Face after upload, then load it with spaCy:

pip install https://huggingface.co/buyanov-tech/ru_ner_v2/resolve/main/ru_ner_v2-any-py3-none-any.whl

import spacy

nlp = spacy.load("ru_ner_v2")
doc = nlp("Труба ПВХ D110 2м серая")

for ent in doc.ents:
    print(ent.text, ent.label_)

Example output from the local model-best checkpoint:

Труба PRODUCT_TYPE
ПВХ SIZE
2м LENGTH
серая COLOR

You can also load the unpacked spaCy pipeline from a local directory:

import spacy

nlp = spacy.load("models/ner-v2/model-best")
doc = nlp("Фильтр воздушный касетный ФВКас-1-590х590х45 G4")
print([(ent.text, ent.label_) for ent in doc.ents])

Example output:

[('Фильтр воздушный касетный', 'PRODUCT_TYPE'), ('ФВКас-1-590х590х45', 'MODEL'), ('G4', 'BRAND')]

Intended Use

This model is intended for internal product catalog processing, product title normalization, feature extraction, and downstream category classification.

Limitations

• The model is trained for Russian product/catalog titles.
• It is optimized for product attribute extraction, not general-purpose NER.
• Label meanings are domain-specific and may differ from generic NLP entity taxonomies.
• Predictions should be validated before use in high-impact catalog updates or automated data enrichment pipelines.