Indic-Bert-NER-BIO-Dataset

A comprehensive Named Entity Recognition dataset in BIO format for training NER models on medical and regulatory documents in Indian languages.

Dataset Overview

This dataset contains annotated medical and pharmaceutical regulatory documents with entity labels in BIO (Begin-Inside-Outside) format. It is designed for training Named Entity Recognition models, particularly the Indic-Bert-NER-Model.

Dataset Statistics

Attribute	Value
Format	BIO (Begin-Inside-Outside) and JSONL
Total Files	7 files (mixed formats)
Languages	Hindi, English, Indian languages
Domain	Medical, Pharmaceutical, Regulatory
Data Sources	CTRI, FAERS, JSL, Phase 2 Augmented

Data Sources

1. CTRI (Clinical Trials Registry - India)

• File: ctri_bio.jsonl (1.1 MB)
• Format: JSONL
• Description: Annotated clinical trial documents from the Indian Clinical Trials Registry
• Content: Drug names, diseases, dosages, routes of administration

2. FAERS (FDA Adverse Event Reporting System)

• File: faers_bio.jsonl (8.3 MB)
• Format: JSONL
• Description: Adverse event reports translated and adapted for Indian pharmaceutical context
• Content: Adverse events, drug names, patient characteristics, outcomes

3. JSL (John Snow Labs Medical Dataset)

• File: jsl_bio.jsonl (410 KB)
• Format: JSONL
• Description: High-quality medical NER annotations from John Snow Labs
• Content: Medical conditions, treatments, medications, procedures

4. Phase 2 Training Data (Augmented)

• Files:
  • phase2_training.bio (15.2 MB)
  • phase2_training_augmented.bio (19.9 MB)
• Format: BIO plain text
• Description: Primary training dataset with augmented samples for robustness
• Content: Comprehensive medical and regulatory entity annotations

5. Phase 2 Merged Dataset

• File: phase2_merged.bio (18.7 MB)
• Format: BIO plain text
• Description: Consolidated dataset combining multiple sources
• Content: De-duplicated and merged annotations

6. Phase 2 Final Augmented

• File: phase2_final_augmented.bio (24.4 MB)
• Format: BIO plain text
• Description: Final augmented dataset with additional synthetic data
• Content: Expanded training set with improved coverage

Entity Schema

Annotated Entity Types

The dataset includes annotations for the following entity categories:

Entity Type	Examples	Use Cases
Drug/Medication	Paracetamol, Aspirin, Ibuprofen	Drug identification in documents
Disease/Condition	Fever, Hypertension, Diabetes	Medical condition extraction
Dosage	500mg, 2 tablets, 1 injection	Medication instructions
Route of Administration	Oral, Intravenous, Subcutaneous	Administration method extraction
Adverse Event	Headache, Rash, Nausea	Safety monitoring
Medical Procedure	Surgery, X-ray, Blood test	Procedure identification
Body Part	Heart, Liver, Kidney	Anatomical references
Regulatory Reference	Schedule H, Form F	Regulatory compliance

BIO Format Explanation

The BIO format works as follows:

• B-LABEL: Beginning of an entity
• I-LABEL: Continuation/Inside of an entity
• O: Outside any entity (not part of an entity)

Example:

Paracetamol    B-Drug
is            O
used          O
for           O
headache      B-Disease
treatment     O
.             O

File Format Details

BIO Format (.bio files)

Plain text with one token per line, entity tags on the same line separated by whitespace:

token1  B-Entity
token2  I-Entity
token3  O

JSONL Format (.jsonl files)

JSON Lines with structured records:

{
  "text": "Paracetamol is used for fever treatment.",
  "entities": [
    {"start": 0, "end": 11, "label": "Drug"},
    {"start": 24, "end": 29, "label": "Disease"}
  ],
  "tokens": ["Paracetamol", "is", "used", "for", "fever", "treatment", "."],
  "tags": ["B-Drug", "O", "O", "O", "B-Disease", "O", "O"]
}

Data Splits

Training Data

• Primary: phase2_training.bio (15.2 MB) - ~45,000 sentences
• Augmented: phase2_training_augmented.bio (19.9 MB) - ~60,000 sentences
• Final: phase2_final_augmented.bio (24.4 MB) - ~70,000 sentences

Validation & Test Data

Embedded in source files (CTRI, FAERS, JSL datasets):

• CTRI: ~500 annotated documents
• FAERS: ~2,000 annotated documents
• JSL: ~300 annotated documents

Data Preparation

Preprocessing Steps

1. Tokenization: Word-level tokenization for BIO format
2. Annotation: Manual and semi-automatic annotation with quality review
3. Augmentation: Synthetic data generation for low-frequency entities
4. Merging: Consolidation of multiple sources with de-duplication
5. Validation: Quality assurance through inter-annotator agreement checks

Known Issues & Limitations

• Some documents contain mixed language text (Hindi + English)
• Entity boundaries may vary between annotators
• Domain-specific terminology variations
• Limited coverage of rare entity types

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset("sharkdodo/Indic-Bert-NER-BIO-Dataset")

# Load specific splits
dataset = load_dataset(
    "sharkdodo/Indic-Bert-NER-BIO-Dataset",
    split="train"
)

# Access individual records
for example in dataset.take(5):
    print(example)

Processing BIO Files

def load_bio_file(filepath):
    """Load BIO format file"""
    sentences = []
    tags = []
    
    current_sent = []
    current_tags = []
    
    with open(filepath, 'r', encoding='utf-8') as f:
        for line in f:
            line = line.strip()
            if not line:  # Empty line = sentence boundary
                if current_sent:
                    sentences.append(current_sent)
                    tags.append(current_tags)
                current_sent = []
                current_tags = []
            else:
                parts = line.split('\t')
                if len(parts) == 2:
                    token, tag = parts
                    current_sent.append(token)
                    current_tags.append(tag)
    
    return sentences, tags

# Usage
sentences, tags = load_bio_file('phase2_training.bio')
print(f"Loaded {len(sentences)} sentences")

Processing JSONL Files

import json

def load_jsonl_file(filepath):
    """Load JSONL format file"""
    data = []
    with open(filepath, 'r', encoding='utf-8') as f:
        for line in f:
            record = json.loads(line)
            data.append(record)
    return data

# Usage
data = load_jsonl_file('ctri_bio.jsonl')
print(f"Loaded {len(data)} records")
for record in data[:3]:
    print(record)

Training with Transformers

from transformers import AutoTokenizer, AutoModelForTokenClassification, Trainer, TrainingArguments
from datasets import load_dataset

# Load dataset
dataset = load_dataset("sharkdodo/Indic-Bert-NER-BIO-Dataset")

# Tokenizer and model
model_name = "ai4bharat/indic-bert"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name, num_labels=num_labels)

# Preprocessing function
def tokenize_and_align_labels(examples):
    tokenized_inputs = tokenizer(
        examples['tokens'],
        truncation=True,
        is_split_into_words=True
    )
    labels = []
    for i, label in enumerate(examples['tags']):
        word_ids = tokenized_inputs.word_ids(batch_index=i)
        label_ids = []
        for word_idx in word_ids:
            if word_idx is None:
                label_ids.append(-100)
            else:
                label_ids.append(label[word_idx])
        labels.append(label_ids)
    tokenized_inputs['labels'] = labels
    return tokenized_inputs

# Tokenize dataset
tokenized_dataset = dataset.map(tokenize_and_align_labels, batched=True)

# Training arguments
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    save_steps=500,
    save_total_limit=2,
    evaluation_strategy="steps"
)

# Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset['train'],
    eval_dataset=tokenized_dataset['validation']
)

trainer.train()

Annotation Guidelines

Quality Standards

• Inter-annotator agreement (IAA) > 0.85 (Cohen's Kappa)
• Double annotation for high-priority documents
• Quality review by domain experts
• Entity boundary agreement protocols

Entity Annotation Rules

1. Drug Names: Include all forms (generic, brand, abbreviations)
2. Diseases: Include conditions and symptoms
3. Dosages: Include numeric value + unit
4. Routes: Standardized terminology
5. Adverse Events: Clinical terms + lay terms

License

This dataset is released under the MIT License.

MIT License

Copyright (c) 2026 Vivek Molleti

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

Privacy & Ethics

Data Privacy

• All personally identifiable information (PII) has been de-identified
• Patient confidentiality protected through anonymization
• Compliance with HIPAA and DPDP guidelines
• No sensitive personal data in the public dataset

Ethical Use

• For medical and regulatory purposes only
• Respect patient privacy at all times
• Do not use for unauthorized commercial purposes
• Support transparency in healthcare AI

Citation

If you use this dataset, please cite:

@dataset{indic_bert_ner_bio_dataset_2026,
  title = {Indic-Bert-NER-BIO-Dataset},
  author = {Vivek Molleti},
  year = {2026},
  url = {https://huggingface.co/datasets/sharkdodo/Indic-Bert-NER-BIO-Dataset},
  note = {Medical and regulatory NER dataset in BIO format for Indian languages}
}

Related Resources

• Trained Model: Indic-Bert-NER-Model
• Base Model: AI4Bharat Indic-BERT

Changelog

Version 1.0 (April 2026)

• Initial release
• 7 dataset files with ~100,000+ sentences
• BIO and JSONL formats
• Multi-source aggregation with quality control

Data Collection Timeline

• Phase 1: CTRI and FAERS data collection (Q1 2025)
• Phase 2: Dataset augmentation and validation (Q3 2025)
• Phase 3: Public release (Q2 2026)