VLURes: Benchmarking Long-Text Grounding and Cross-Lingual Robustness in Vision Language Models

Dataset Description

VLURes is a multilingual benchmark for evaluating the fine-grained visual and linguistic understanding of Vision-Language Models (VLMs) in long-text settings. It was created to move beyond short-caption, English-centric evaluation and instead test image understanding, long-context grounding, and cross-lingual robustness in culturally diverse settings.

This dataset is associated with our ACL2026 Findings paper titled "VLURes: Benchmarking Long-Text Grounding and Cross-Lingual Robustness in Vision Language Models."

The current Hugging Face release contains the uploaded image-text pairs in a single multilingual split, with each example consisting of a renamed image file, its paired long-form text, and a language identifier.

Key Features

• Multilingual and culturally grounded: The dataset covers English, Japanese, Swahili, and Urdu.
• Long-text grounding: Each example pairs an image with substantially richer text than standard short-caption benchmarks.
• Single multilingual release: The uploaded Hugging Face version is organized as one train split, with language identified in the language field.
• Benchmark-oriented design: The data supports fine-grained evaluation of VLMs across visual, linguistic, and cross-modal tasks.
• Low-resource language coverage: The benchmark includes dedicated resources for Swahili and Urdu, which remain underrepresented in existing vision-language evaluation datasets.

Supported Tasks

VLURes was designed to support evaluation across eight tasks:

1. Object Recognition (OR)
2. Scene Understanding (SU)
3. Relationship Understanding (RU)
4. Semantic Segmentation (SS)
5. Image Captioning (IC)
6. Image-Text Matching (ITM)
7. Unrelatedness (U)
8. Visual Question Answering (VQA)

The Hugging Face release provides the core multilingual image-text pairs. Task prompts, evaluation protocols, and benchmark-specific task formulations are described in the paper and accompanying project materials.

---

Dataset Structure

Repository Layout

The uploaded dataset follows the structure below:

VLURes_hf_ready/
├── README.md
└── train/
    ├── metadata.parquet
    └── images/
        ├── en/...
        ├── sw/...
        ├── ur/...
        └── jp/...

Data Format

The dataset is packaged in an ImageFolder-style format for easy loading with the Hugging Face datasets library.

• train/metadata.parquet stores the metadata table.
• train/images/... contains the actual image files.
• All images in this release are stored as .jpg files after preprocessing.

Data Instances

In the uploaded metadata.parquet, each row contains the following fields:

• id: a unique example identifier such as en_000001
• file_name: the relative path to the image file, for example images/en/en_000001.jpg
• text: the paired text associated with the image
• language: the language code for the example

When loaded with the Hugging Face datasets library, the dataset exposes the following features:

• id
• image
• text
• language

Language Codes

The current uploaded release uses the following values in the language field:

• en for English
• jp for Japanese
• sw for Swahili
• ur for Urdu

Note that the metadata uses jp in the actual uploaded files for compatibility with the current release structure.

Splits

This Hugging Face release currently provides a single split:

• train

This split contains the multilingual image-text pairs used in the benchmark release.

Data Size

The current uploaded release contains 3,415 examples in total.

Language	Number of image-text pairs
English (en)	996
Swahili (sw)	1,030
Urdu (ur)	949
Japanese (jp)	440
Total	3,415

We have not included lots of Japanese (ja) image-text pairs in this release due to license restrictions imposed by the respective web sources. For en, sw, ur, we have removed some image-text pairs as well.

---

Dataset Creation

Curation Rationale

VLURes was created to evaluate VLMs in settings that require more than shallow image-caption matching. The benchmark emphasizes:

1. multilingual understanding,
2. culturally grounded content,
3. long-text visual grounding,
4. robustness beyond English-only evaluation, and
5. fine-grained multimodal reasoning.

Source Data

The image-text pairs were curated from publicly accessible web sources, including encyclopedia-style pages, news content, and other article-like web documents containing naturally co-occurring images and text.

The benchmark spans a broad range of topics, including:

• animals
• products
• buildings
• locations
• events
• food
• drinks
• hobbies
• works of art
• organizations

Image-Text Alignment

For each document, candidate images were matched to the article content and filtered to retain representative image-text pairs suitable for benchmark construction. The final release stores the prepared image files together with their corresponding text in a format that is easy to load and use for research.

Preprocessing for the Hugging Face Release

For this uploaded release:

• image files were converted into a unified .jpg format,
• files were renamed into stable identifiers such as en_000001.jpg,
• text content was extracted and cleaned from source text files,
• the dataset was organized into a single multilingual train split,
• metadata was consolidated into metadata.parquet.

---

Usage

You can load the dataset directly from Hugging Face using the datasets library:

from datasets import load_dataset

dataset = load_dataset("atamiles/VLURes")
print(dataset)
print(dataset["train"][0])

A typical example will contain:

{
    "id": "en_000001",
    "image": <PIL image>,
    "text": "...",
    "language": "en"
}

To inspect the language distribution:

from collections import Counter

langs = Counter(dataset["train"]["language"])
print(langs)

---

Intended Uses

VLURes is intended for research use in:

• multilingual vision-language evaluation,
• long-text visual grounding,
• cross-lingual robustness analysis,
• multimodal benchmarking,
• low-resource language research.

It may also be useful for studying failure modes of VLMs under long-context and multilingual conditions.

Out-of-Scope Uses

This dataset was not designed for:

• face recognition or identity inference,
• surveillance applications,
• safety-critical deployment without additional validation,
• legal or medical decision-making,
• commercial reuse without checking the rights associated with the underlying source materials.

---

Important Note on Copyright and Licensing

The benchmark release is shared for research use under the license specified for this repository.

Because the data originates from public web sources, users are responsible for ensuring that their use of the released materials complies with any applicable third-party rights, copyright restrictions, and terms of use associated with the original source content.

If you plan to redistribute, adapt, or deploy the contents beyond research use, please verify the status of the original source materials independently.

---

Citation

If you use VLURes in your work, please cite the associated paper:

@inproceedings{atuhurra-etal-2026-vlures,
    title = "{VLUR}es: Benchmarking Long-Text Grounding and Cross-Lingual Robustness in Vision Language Models",
    author = "Atuhurra, Jesse  and
      Ali, Iqra  and
      Iwakura, Tomoya  and
      Kamigaito, Hidetaka  and
      Hiraoka, Tatsuya",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1367/",
    pages = "27426--27481",
    ISBN = "979-8-89176-395-1",
    abstract = "We introduce ***VLURes***, a multilingual benchmark for evaluating Vision-Language Models (VLMs) under *long-text grounding*: selecting and reasoning over the image-relevant subset of article-length text that contains distractors and ungrounded claims. *VLURes* contains **4,000** web-curated *image + long-text* pairs across **English (En), Japanese (Ja), Swahili (Sw), and Urdu (Ur)** and **10** topical categories, and defines **eight** tasks spanning image-only perception (OR, SU, RU, SS, IC) and image+text grounding (ITM, *Unrelatedness*, VQA). To construct web-realistic pairs, we apply language-adapted CLIP alignment to select representative images and filter weakly grounded pages. Across **10** proprietary and open VLMs evaluated under zero-shot and one-shot prompting, with and without rationales, the best model (GPT-4o) reaches **90.8{\%}** overall accuracy but remains **6.7** points below human performance (**97.5{\%}**) on Object Recognition, and cross-lingual sensitivity persists, while open models are substantially weaker and often lack reliable multilingual VL support. *VLURes* provides a practical testbed for long-text grounding and multilingual robustness in web-realistic agent settings."
}

---

Contact

For questions about the dataset, benchmark, or associated paper, please use the project repository or contact Jesse Atuhurra.