Datasets:
license: mit
dataset_info:
features:
- name: tokens
list: string
- name: ner_tags
list: string
splits:
- name: train
num_bytes: 1226892
num_examples: 850
download_size: 325452
dataset_size: 1226892
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- token-classification
language:
- ru
- en
pretty_name: Jay Guard NER Benchmark
Jay Guard NER Benchmark
Dataset Description
The Jay Guard NER Benchmark is a Russian-language dataset designed for evaluating Named Entity Recognition (NER) models on their ability to identify personal and sensitive data. The data is sourced from real-world, complex conversational texts, including work chats, customer support logs, and spoken language transcripts.
This dataset was created by Just AI to benchmark the performance of various NER solutions for the task of data anonymization. It specifically targets entities that are critical for protecting personal data, such as PERSON and STREET_ADDRESS.
Supported Tasks and Leaderboards
The dataset is intended for the Named Entity Recognition (NER) task. The goal is to train and evaluate models that can accurately identify and classify tokens into predefined categories.
Languages
The text in the dataset is in Russian (ru). It contains both Cyrillic and Latin characters, reflecting real-world usage in chats and logs.
Dataset Structure
The dataset consists of a single configuration and is split into train, validation, and test sets.
Data Instances
Each instance in the dataset consists of a list of tokens and a corresponding list of ner_tags.
An example from the dataset looks like this:
{
"tokens": ["Слушай", ",", "я", "в", "2005", "в", "Москве", "на", "Тверской", "15", "стави", "..."],
"ner_tags":
}
Data Splits
The data is split into:
- train: 850 examples
Dataset Creation
Curation Rationale
The dataset was created to address the lack of robust benchmarks for personal data detection in complex, noisy Russian conversational text. Standard NER models often fail in these scenarios, either by missing entities (low recall) or by incorrectly tagging non-sensitive information (low precision). This benchmark serves to evaluate and improve models for real-world data anonymization tasks.
Source Data
The source data is derived from internal, anonymized logs of conversational systems, work chats, and customer support interactions. All data has been processed to remove or replace any real personally identifiable information (PII) before being included in this public benchmark.
Considerations for Using the Data
Social Impact of Dataset
This dataset is designed to facilitate the development of more effective and reliable data anonymization technologies. By providing a challenging benchmark, we aim to help researchers and developers build NLP systems that more effectively protect user privacy.
Discussion of Biases
While the data is sourced from a variety of conversational contexts, it may reflect the linguistic patterns and biases present in those sources. The data primarily comes from Russian-speaking users in specific technical or customer-support domains, and models trained on this data may not generalize perfectly to other domains (e.g., legal or medical texts).
Other Known Limitations
The dataset focuses primarily on PERSON (PERSON, PUBLIC_PERSON, FICT), GPE, and STREET_ADDRESS entities. It does not cover other types of PII, such as phone numbers, email addresses, or financial information, which would require separate detection models.
Additional Information
Dataset Curators
This dataset was curated by the team at Just AI as part of the development of the Jay Guard.
Licensing Information
The dataset is licensed under the Apache License, Version 2.0.
Citation Information
If you use this dataset in your research, please cite it as follows:
@misc{jayguard_ner_benchmark,
author = {Just AI},
title = {Jay Guard NER Benchmark},
year = {2025},
publisher = {Hugging Face},
journal = {Hugging Face Datasets},
url = {https://huggingface.co/datasets/just-ai/jayguard-ner-benchmark}
}