Datasets:

AIIRLab
/

MATVB

configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - MATVB_train.json
          - MATVB_query_nl.json
          - MATVB_query_caption.json
task_categories:
  - image-to-text
  - image-retrieval
language:
  - en
tags:
  - image-captioning
  - image-text-retrieval
  - mathematics
  - multimodal
  - benchmark
  - wikipedia
size_categories:
  - 1K<n<10K

MAT-VB: MAthematical Text-Vision Benchmark

🏆 Best Resource Paper Award — 2025 ACM/IEEE Joint Conference on Digital Libraries (JCDL)

Official Hugging Face repository for MAT-VB, a multimodal benchmark constructed from Wikipedia mathematical content designed to evaluate mathematical image captioning and text-to-image retrieval models.

🔗 Links

GitHub Repository: AIIRLab/MATVB
Paper (IEEE Xplore): MAT-VB: MAthematical Text-Vision Benchmark

📁 Dataset Files Overview

File Name	Format	Description
`images.zip`	Archive (`.zip`)	Contains all ~4,000 Wikipedia math images (216 MB uncompressed).
`MATVB_train.json`	JSON	Training annotations. Maps `image_id` (key) to its ground-truth `caption` (value).
`MATVB_query_nl.json`	JSON	Natural language query set for text-to-image search (human-authored intent queries).
`MATVB_query_caption.json`	JSON	Caption query set for text-to-image search (original Wikipedia captions).
`MATVB_qrel.qrel`	TREC format	Evaluation relevance file mapping query IDs to target image IDs.

🎯 Task Definitions & Splits

1. Mathematical Image Captioning

Training Set: Use all entries in MATVB_train.json (3,600 image-caption pairs).
Test Set: Any image in images.zip whose ID is not present in MATVB_train.json serves as a test sample (400 test images). Models generate captions for these test images during evaluation.

2. Text-to-Image Retrieval

The benchmark provides 400 test evaluation queries under two modalities sharing identical query IDs:

Natural Language Queries (MATVB_query_nl.json): Human-authored search queries representing intent-based user searches.
Caption Queries (MATVB_query_caption.json): Queries built using original, detailed Wikipedia captions.
Ground Truth (MATVB_qrel.qrel): Use standard TREC evaluation tools (e.g., trec_eval) to score retrieval runs against this relevance file.

🚀 How to Load and Use

You can download and parse the dataset files directly using huggingface_hub and standard Python libraries:

import json
import zipfile
from huggingface_hub import hf_hub_download

# 1. Download dataset files
repo_id = "AIIRLab/MATVB"

train_path = hf_hub_download(repo_id=repo_id, filename="MATVB_train.json", repo_type="dataset")
query_nl_path = hf_hub_download(repo_id=repo_id, filename="MATVB_query_nl.json", repo_type="dataset")
query_cap_path = hf_hub_download(repo_id=repo_id, filename="MATVB_query_caption.json", repo_type="dataset")
qrel_path = hf_hub_download(repo_id=repo_id, filename="MATVB_qrel.qrel", repo_type="dataset")
images_zip_path = hf_hub_download(repo_id=repo_id, filename="images.zip", repo_type="dataset")

# 2. Extract image archive
with zipfile.ZipFile(images_zip_path, 'r') as zip_ref:
    zip_ref.extractall("./images")

# 3. Load JSON annotations
with open(train_path, "r") as f:
    train_data = json.load(f)  # dict: {image_id: caption}

with open(query_nl_path, "r") as f:
    queries_nl = json.load(f)  # dict: {query_id: query_text}

with open(query_cap_path, "r") as f:
    queries_caption = json.load(f)  # dict: {query_id: query_text}

print(f"Loaded {len(train_data)} training samples.")
print(f"Loaded {len(queries_nl)} NL retrieval queries.")