---
dataset_info:
  features:
  - name: input_ids
    sequence: int32
  - name: attention_mask
    sequence: int8
  - name: labels
    sequence: int64
  splits:
  - name: train
    num_bytes: 76440290.15811698
    num_examples: 120113
  - name: test
    num_bytes: 19110549.84188302
    num_examples: 30029
  download_size: 16997872
  dataset_size: 95550840.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

## dataset summary
NER-News-BIDataset is a dataset for named entity recognition (NER) in news articles, publicly released by the National Institute of Korean Language in 2023.   
The dataset is labeled with named entities specifically for news data.   
It consists of a total of 150,142 sentences, and entities are categorized into 150 labels for recognition.  

## Languages
korean  

## Data Structure
DatasetDict({
    train: Dataset({
        features: ['input_ids', 'attention_mask', 'labels'],
        num_rows: 120113
    })
    test: Dataset({
        features: ['input_ids', 'attention_mask', 'labels'],
        num_rows: 30029
    })
})  

### Data Instances
The dataset is provided in text format with train/test sets.   
Each instance represents a news article, and if there is an entity in the sentence, it is appropriately tagged with the corresponding label.   
In cases where a single entity is separated into multiple tokens, the first token is labeled as "B-entity" and the subsequent tokens are labeled as "I-entity" until the end.  

### Data Fields