README.md

---
license: cc-by-4.0
task_categories:
  - text-generation
  - summarization
  - image-to-text
language:
  - bn
tags:
  - bengali
  - news
  - headline-generation
  - multimodal
  - vision-language
  - nlp
size_categories:
  - 1M<n<10M
---

# BartaLens: Bengali Multimodal News Corpus

<p align="center">
  <a href="https://arxiv.org/abs/0000.00000">
    <img src="https://img.shields.io/badge/%F0%9F%93%84_Paper-Coming_Soon-b12a00?style=for-the-badge&labelColor=ffb300" alt="Paper Coming Soon">
  </a>
</p>

[![Paper](https://img.shields.io/badge/arXiv-coming--soon-red)](https://arxiv.org/abs/0000.00000)
[![GitHub](https://img.shields.io/badge/GitHub-Code-blue)](https://github.com/dipta007/bengali-news-headline)
[![Text-Only Version](https://img.shields.io/badge/HuggingFace-Barta_(text--only)-yellow)](https://huggingface.co/datasets/dipta007/Barta)

## Overview

**BartaLens** (বার্তা + Lens) is a large-scale multimodal Bengali news corpus pairing articles with their associated news images. It is designed for training and evaluating vision-language models on Bengali headline generation.

Key features:
- **1.2M+ articles** with associated news images
- **Three splits** (train / validation / test) for reproducible evaluation
- **Multimodal** — each article paired with its news photograph
- **Diverse sources** from major Bangladeshi news outlets

## Dataset Statistics

| Split | Rows |
|-------|-----:|
| train | 1,219,442 |
| validation | 5,000 |
| test | 15,000 |

## Loading the Dataset

```python
from datasets import load_dataset

# Load train split
dataset = load_dataset("dipta007/BartaLens", split="train")

# Load test split
test = load_dataset("dipta007/BartaLens", split="test")

# Access an example
example = dataset[0]
image = example["image"]  # PIL Image
headline = example["headline"]
article = example["article"]
```

## Data Fields

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique identifier |
| `article` | string | Full article body in Bengali |
| `headline` | string | Article headline in Bengali |
| `source` | string | News outlet name |
| `category` | string | News category in Bengali |
| `image` | Image | Associated news photograph (PIL Image) |

## Intended Use

- **Multimodal headline generation** — generate headlines from article text + image
- **Vision-language model** fine-tuning (e.g., Gemma, Qwen VLMs with LoRA)
- **Image-guided summarization** — leverage visual context for better summaries
- **Bengali VLM benchmarking** — evaluate multimodal understanding in Bengali
- **Cross-modal retrieval** — match news images to articles

## Related Datasets

- [**Barta**](https://huggingface.co/datasets/dipta007/Barta) — text-only version (1.7M articles, full corpus without image filtering)

## Citation

```bibtex
@article{BartaLens2025,
  title={BartaLens: A Multimodal Bengali News Corpus for Headline Generation},
  author={Shubhashis Roy Dipta},
  year={2025}
}
```