---
language:
  - bn
  - en
tags:
  - bangla
  - bengali
  - nlp
  - text-corpus
  - pretraining
  - social-media
  - news
  - sentiment
  - bangla-nlp
  - bengali-nlp
  - bangla-dataset
  - bengali-dataset
  - bangla-text
  - bangla-llm
  - bangladesh
license: mit
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.jsonl
---

# Bangla NLP Text Corpus — 800K+ Bangla Text Samples for LLM Training and NLP Research

The largest open, multi-domain Bangla text dataset, combining **801,645 samples** from 15 different sources — newspapers, social media, education, reviews, QA, medical, poetry, and more. Ready for Bangla LLM pretraining, fine-tuning, and downstream NLP tasks.

## Why This Dataset?

Bangla (Bengali) is the 7th most spoken language in the world with 230M+ speakers, but high-quality Bangla NLP datasets are scarce and scattered. This collection brings together cleaned, deduplicated text from diverse domains into a single, easy-to-use dataset with a consistent schema.

Good for:
- **Bangla LLM pretraining** and continued pretraining
- **Bangla text classification** (sentiment, emotion, sarcasm)
- **Bangla language modeling** and perplexity evaluation
- **Bangla NER, POS tagging**, and other token-level tasks
- **Bangla RAG pipelines** and semantic search indices
- **Bangla text generation** fine-tuning
- **Cross-lingual transfer learning** (Bengali-English)

## Dataset Overview

| Source | Samples | Domain | Description |
|--------|---------|--------|-------------|
| potrika_newspaper | 326,988 | News | Bangla newspaper articles from Potrika |
| addressacademy | 292,919 | General | Diverse Bangla text corpus |
| facebook_sentiment | 40,146 | Social Media | Facebook posts with sentiment |
| dataset_csv | 35,421 | Mixed | Assorted Bangla text |
| bangla_sarcasm | 23,592 | Social Media | Sarcasm-labeled Bangla text |
| bangla_emotion | 19,371 | Social Media | Emotion-labeled Bangla text |
| restaurant_reviews | 15,684 | Reviews | Bangla restaurant reviews |
| youtube_sentiment | 11,783 | Social Media | YouTube comment sentiment |
| nctb | 11,257 | Education | NCTB Bangladesh textbook content |
| tydiqa_bengali | 10,768 | QA | Bengali question-answering pairs |
| bangla_telecom | 6,546 | Telecom | Telecom domain Bangla text |
| reddit | 5,605 | Social Media | Reddit Bangla posts |
| deshimula | 897 | Reviews | Deshimula Bangla reviews |
| bengali_medical | 658 | Medical | Bengali medical text |
| bengali_poetry | 10 | Literature | Bengali poetry |

**Total: 801,645 samples across 15 sources**

## Schema

Every row follows the same JSON structure:

```json
{
  "text": "বাংলাদেশের অর্থনীতি দ্রুত বৃদ্ধি পাচ্ছে...",
  "source": "potrika_newspaper",
  "meta": {}
}
```

| Field | Type | Description |
|-------|------|-------------|
| `text` | string | The Bangla text content |
| `source` | string | Origin dataset/domain (see table above) |
| `meta` | object | Optional metadata (labels, categories, etc.) |

## Quick Start

```python
from datasets import load_dataset

# Load full dataset
ds = load_dataset("nahidstaq/bangla-llm-data")
print(f"Total samples: {len(ds['train'])}")  # 801,645

# Filter by domain
news = ds["train"].filter(lambda x: x["source"] == "potrika_newspaper")
sentiment = ds["train"].filter(lambda x: x["source"] in ["facebook_sentiment", "youtube_sentiment"])
education = ds["train"].filter(lambda x: x["source"] == "nctb")

# Get all unique sources
sources = set(ds["train"]["source"])
print(sources)
```

## Sample Texts

**News** (potrika_newspaper):
> রাজধানী ঢাকায় আজ সকাল থেকেই ভারি বৃষ্টিপাত হচ্ছে। আবহাওয়া অধিদপ্তর জানিয়েছে...

**Social Media** (facebook_sentiment):
> ভাই এই ফোনটা কিনে ভুল করেছি, ক্যামেরা একদম খারাপ 😡

**Education** (nctb):
> বাংলাদেশ দক্ষিণ এশিয়ার একটি দেশ। এর রাজধানী ঢাকা। বাংলাদেশের মোট আয়তন...

**Restaurant Review** (restaurant_reviews):
> খাবারের মান ভালো ছিল কিন্তু সার্ভিস একটু ধীর গতির

**Medical** (bengali_medical):
> ডায়াবেটিস রোগীদের নিয়মিত রক্তে শর্করার মাত্রা পরীক্ষা করা উচিত

## Related Resources

- [nahidstaq/bangla-transliteration](https://huggingface.co/nahidstaq/bangla-transliteration) — Roman to Bangla transliteration model (9.18% CER)
- [nahidstaq/bge-m3-bangla](https://huggingface.co/nahidstaq/bge-m3-bangla) — Cross-script Bangla semantic search model
- [nahidstaq/bangla-transliteration-data](https://huggingface.co/datasets/nahidstaq/bangla-transliteration-data) — Bangla transliteration pairs

## Credits

Collected and curated by **Md Nahid Hasan** ([@nahidstaq](https://huggingface.co/nahidstaq) | [nahidscript.com](https://nahidscript.com)).

Individual sources retain their original licenses. This collection is released under MIT for the curation and aggregation work.

## Citation

```bibtex
@misc{bangla-llm-data,
  author = {Hasan, Md Nahid},
  title = {Bangla LLM Data: 800K+ Multi-Domain Bangla Text Corpus},
  year = {2025},
  url = {https://huggingface.co/datasets/nahidstaq/bangla-llm-data},
  note = {Curated multi-source Bangla text dataset for LLM training and NLP research}
}
```