---
pretty_name: Punjabi Multimodal Visual Reasoning (PuMVR)
tags:
- multimodal
- visual-question-answering
- multi-script
- low-resource-language
- punjabi
- image-to-text
- multiple-choice
language:
- pa
- en
language_bcp47:
- pa-Guru
- pa-Arab
- pa-Latn
task_categories:
- visual-question-answering
- image-to-text
- multiple-choice
- question-answering
license: cc-by-4.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
multilinguality: multi-script
annotations_creators:
- human
language_creators:
- native-speakers
size_categories:
- 100M<X<1B
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: reasoning
    dtype: string
  - name: scripts_gurmukhi_question
    dtype: string
  - name: scripts_gurmukhi_options
    list: string
  - name: scripts_gurmukhi_answer
    dtype: string
  - name: scripts_shahmukhi_question
    dtype: string
  - name: scripts_shahmukhi_options
    list: string
  - name: scripts_shahmukhi_answer
    dtype: string
  - name: scripts_roman_question
    dtype: string
  - name: scripts_roman_options
    list: string
  - name: scripts_roman_answer
    dtype: string
  splits:
  - name: train
    num_bytes: 784495831
    num_examples: 1000
  download_size: 1568664333
  dataset_size: 784495831
---

# PuMVR: Punjabi Multimodal Visual Reasoning Benchmark

## Paper
This dataset accompanies the paper: [PuMVR](https://huggingface.co/papers/2606.17188)

## 🌟 Dataset Overview

**PuMVR (Punjabi Multimodal Visual Reasoning)** is a novel benchmark designed to evaluate **script-dependent performance biases** in Vision-Language Models (VLMs). It addresses the critical gap that current VLM evaluations fail to test whether models are truly **multi-script**, a distinction vital for languages like Punjabi which are actively written in multiple scripts.

The dataset features **1000 unique image-text reasoning tasks** focused on Punjabi culture, history, and daily life. All instances are translated and rigorously validated across the three active Punjabi writing systems: **Gurmukhi (pa-Guru), Shahmukhi (pa-Arab), and Roman (pa-Latn)**.

* **Total Instances:** 1000
* **Total Size:** 749 MB
* **Language:** Punjabi (pa) with three distinct script variants.
* **Target Models:** State-of-the-art VLMs
  
---

## 📊 Dataset Structure and Statistics

The dataset is organized into a single split (`train`) and is composed of image data and corresponding textual annotations stored in a JSON file.

### Data Fields

The dataset schema contains all necessary components for running multiple-choice VQA across three scripts:

| Field Name | Data Type | Description |
| :--- | :--- | :--- |
| **`id`** | `string` | Unique identifier (e.g., `I_1.png`). |
| **`image`** | `Image` | The associated visual input (decoded from the file path). |
| **`reasoning`** | `string` | Human-written explanation for the ground truth answer (in English). |
| **`scripts_[script]_question`** | `string` | The reasoning question in the specified script. |
| **`scripts_[script]_options`** | `list[string]`| 4 multiple-choice options in the specified script. |
| **`scripts_[script]_answer`** | `string` | The single correct option in the specified script. |

*(The `[script]` placeholder is one of: `gurmukhi`, `shahmukhi`, or `roman`.)*

---

## ⚖️ Ethical and Legal Considerations

### Licenses
* **Data:** The PuMVR dataset is released under the **Creative Commons Attribution 4.0 International License (CC BY 4.0)**.
* **Images:** Majority (approximately 95%) of the images are AI-generated (synthetic data) to ensure cultural specificity and clear licensing. The remaining images are sourced from public domain, Wikimedia Commons, and original photography. 

### Data Creation and Validation
The textual data was created and rigorously validated by a team of native speakers across both Indian and Pakistani Punjabi contexts to ensure **semantic equivalence** and **cultural appropriateness** across the Gurmukhi, Shahmukhi, and Roman scripts.

### Limitations
The dataset is highly focused on **Punjabi culture**, which introduces a domain-specific bias. The Romanization used reflects common digital usage but is not strictly standardized, mirroring real-world multi-script challenges.