Modelpublic

Martsv07/rubert-ner-search-queries

sourceHugging Faceapache-2.0updated 11mo agoView on Hugging Face

0likes26downloads

Model Card

RuBERT NER for Russian Search Queries

This model is a fine-tuned version of `DeepPavlov/rubert-base-cased` for Named Entity Recognition (NER) on short and noisy Russian search queries typical of grocery and e-commerce platforms.

The model identifies key product-related entities using the BIO tagging scheme with the following labels:

—TYPE — product type or category (e.g., молоко, чипсы, вода)
—BRAND — brand name (e.g., Lay’s, Простоквашино)
—VOLUME — quantity or size (e.g., 0.5 л, 200 г, 10 шт)
—PERCENT — percentage or fat content (e.g., 2.5%, 15%)

🧠 Model Description

—Architecture: BERT (RuBERT cased)
—Task: Token classification / Named Entity Recognition
—Language: Russian
—Tagging scheme: BIO
—Entity types: TYPE, BRAND, VOLUME, PERCENT
—Training data: Private hackathon dataset from X5 Group Company of anonymized Russian search queries (not publicly available)

The model is optimized for short, informal queries and handles casing variation, abbreviations, and mild typos.

🚀 Usage

This repository contains only the fine-tuned model weights and tokenizer configuration. A minimal example of how to load the model:

python

from transformers import AutoTokenizer, AutoModelForTokenClassification

model_id = "Martsv07/rubert-ner-search-queries"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

⚖️ License

Distributed under the Apache-2.0 license, in accordance with the license of the base model `DeepPavlov/rubert-base-cased`.

👥 Authors

Developed by @Martsv07.

🔗 Related Repository

For the full inference pipeline — including preprocessing, postprocessing, and FastAPI service implementation — see the accompanying GitHub repository.