abhilash88/odia-text-corpus

Odia Text Corpus

Dataset Description

This is a comprehensive Odia language text corpus designed for training language models, text generation, and various NLP tasks in Odia (ଓଡ଼ିଆ). The dataset contains high-quality Odia text from multiple sources, providing a rich foundation for Odia language AI development.

Dataset Summary

• Language: Odia (ଓଡ଼ିଆ)
• Total Records: 649,120
• Text Format: Plain Odia text
• License: CC-BY-4.0
• Use Cases: Language modeling, text generation, classification, analysis

Dataset Structure

Data Splits

The dataset is organized into the following splits:

• train: 292,104 records - Primary training data
• validation: 32,456 records - Validation data
• mistral_train: 324,560 records - Additional training data

Data Fields

Each record contains:

• id: Unique identifier for the text
• text: The main Odia text content
• char_count: Number of characters in the text
• word_count: Number of words in the text
• source_file: Original source file name
• dataset_type: Type classification of the data
• format: Data format identifier

Example Record

{
  "id": "train_12345",
  "text": "ଓଡ଼ିଶାର ସାଂସ୍କୃତିକ ପରମ୍ପରା ବହୁତ ସମୃଦ୍ଧ ଅଟେ। ଏଠାରେ ଅନେକ ପର୍ବପର୍ବାଣୀ ପାଳିତ ହୁଏ...",
  "char_count": 1258,
  "word_count": 185,
  "source_file": "odia_train.jsonl",
  "dataset_type": "train_text",
  "format": "text"
}

Supported Tasks

Primary Tasks

• Language Modeling: Train Odia language models (GPT, BERT, etc.)
• Text Generation: Generate coherent Odia text
• Masked Language Modeling: Fill-in-the-blank tasks in Odia
• Text Classification: Classify Odia text by topic/sentiment

Secondary Tasks

• Text Summarization: Generate summaries of Odia text
• Question Answering: Build Odia QA systems
• Machine Translation: Odia-English translation models
• Sentiment Analysis: Analyze sentiment in Odia text

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("abhilash88/odia-text-corpus")

# Load specific splits
train_data = dataset["train"]
validation_data = dataset["validation"]
mistral_data = dataset["mistral_train"]

print(f"Train: {len(train_data):,} records")
print(f"Validation: {len(validation_data):,} records")
print(f"Mistral: {len(mistral_data):,} records")

Training a Language Model

from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer

# Load tokenizer (use an Indic language tokenizer)
tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-bert")

# Tokenize the dataset
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding=True, max_length=512)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

# Initialize model
model = AutoModelForCausalLM.from_pretrained("gpt2")  # or your preferred model

# Set up training
training_args = TrainingArguments(
    output_dir="./odia-language-model",
    overwrite_output_dir=True,
    num_train_epochs=3,
    per_device_train_batch_size=4,
    save_steps=10_000,
    save_total_limit=2,
)

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

# Train the model
trainer.train()

Text Analysis

# Analyze text characteristics
import pandas as pd

# Convert to pandas for analysis
df = dataset["train"].to_pandas()

# Basic statistics
print(f"Average text length: {df['char_count'].mean():.0f} characters")
print(f"Average words per text: {df['word_count'].mean():.0f} words")
print(f"Longest text: {df['char_count'].max():,} characters")
print(f"Shortest text: {df['char_count'].min():,} characters")

# Text length distribution
import matplotlib.pyplot as plt
df['char_count'].hist(bins=50)
plt.xlabel('Text Length (characters)')
plt.ylabel('Frequency')
plt.title('Distribution of Text Lengths')
plt.show()

Data Quality and Preprocessing

Quality Assurance

• ✅ Language Validation: All text verified as Odia script
• ✅ Encoding: Proper UTF-8 encoding for Odia characters
• ✅ Deduplication: Duplicate removal based on content
• ✅ Length Filtering: Removed extremely short or empty texts
• ✅ Format Consistency: Standardized structure across all records

Preprocessing Steps

1. Text Extraction: Extracted from original JSON/JSONL files
2. Character Encoding: Ensured proper Odia Unicode support
3. Length Calculation: Added character and word counts
4. Metadata Addition: Added source and type information
5. Quality Filtering: Removed low-quality or corrupted text

Dataset Statistics

Overall Statistics

• Total Records: 649,120
• Total Characters: ~649,120,000 (estimated)
• Total Words: ~97,368,000 (estimated)
• Average Text Length: ~1000 characters
• Language Coverage: Modern Odia (ଓଡ଼ିଆ)

Split Distribution

• Training Data: 616,664 records (95.0%)
• Validation Data: 32,456 records (5.0%)

Ethical Considerations

Responsible Use

• Attribution: Maintain proper attribution to original sources
• Academic Use: Primarily intended for research and education
• Commercial Use: Contact dataset creator for commercial applications
• Cultural Sensitivity: Respect Odia cultural context and values

Limitations

• Temporal Coverage: Data from specific time periods
• Domain Coverage: May not represent all Odia text varieties
• Regional Dialects: Primarily standard Odia, limited dialect coverage
• Bias Considerations: May reflect biases present in source data

Recommended Use Cases

• ✅ Academic research in Odia NLP
• ✅ Educational applications and tools
• ✅ Language preservation efforts
• ✅ AI model development for Odia
• ❌ Commercial use without proper licensing
• ❌ Applications that may harm Odia language or culture

Technical Specifications

File Format

• Storage: Parquet format for efficient loading
• Encoding: UTF-8 with proper Odia Unicode support
• Compression: Optimized for size and speed
• Compatibility: Works with pandas, datasets, and other tools

Performance

• Loading Speed: Optimized for fast data loading
• Memory Usage: Efficient memory utilization
• Processing: Suitable for large-scale model training
• Streaming: Supports streaming for large datasets

Citation

If you use this dataset in your research, please cite:

@dataset{odia_text_corpus_2025,
  title={Odia Text Corpus},
  author={Abhilash},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/abhilash88/odia-text-corpus}
}

Acknowledgments

• Odia Language Community: For preserving and promoting Odia
• Data Contributors: Original text creators and sources
• Open Source Community: For tools and infrastructure
• Hugging Face: For hosting and distribution platform

Contact and Support

• Dataset Creator: abhilash88
• Issues: Please report issues via the dataset repository
• Contributions: Community contributions welcome
• Updates: Regular updates planned based on community feedback

Future Work

• Expansion: Addition of more Odia text sources
• Quality Improvement: Enhanced preprocessing and filtering
• Dialect Coverage: Inclusion of Odia regional dialects
• Domain Expansion: Coverage of specialized domains (technical, legal, etc.)
• Community Contributions: Integration of community-provided data

---

This dataset is part of efforts to advance Odia language technology and promote digital inclusion for Odia speakers worldwide. 🚀

Keywords: Odia, ଓଡ଼ିଆ, Indian Languages, NLP, Language Modeling, Text Generation, Low-Resource Languages