README.md

---
task_categories:
- image-to-text
viewer: false
---
# RAVENEA

[**📃 Paper**](https://arxiv.org/abs/2505.14462) | [**🌐 Project Page**](https://jiaangli.github.io/ravenea/) | [**💻 Github**](https://github.com/yfyuan01/RAVENEA)

**RAVENEA** is a multimodal benchmark designed to comprehensively evaluate the capabilities of VLMs in **cultural understanding through RAG**, introduced in [RAVENEA: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding](https://arxiv.org/abs/2505.14462).

It provides:

- **A large-scale cultural retrieval-generation corpus** featuring 1,868 culturally grounded images paired with over 10,000 **human-ranked** Wikipedia documents.

- **Two downstream tasks** for assessing culture-centric visual understanding (cVQA) and culture-informed image captioning (cIC).

- **Broad cross-cultural coverage spanning 8 countries and 11 categories**, including China, India, Indonesia, Korea, Mexico, Nigeria, Russia, and Spain. The benchmark encompasses a diverse taxonomic spectrum: Architecture, Cuisine, History, Art, Daily Life, Companies, Sports & Recreation, Transportation, Religion, Nature, and Tools.


## Dataset Structure

The dataset is organized as follows:

```
ravenea/
├── images/                  # Directory containing all images
├── metadata_train.jsonl     # Training split metadata
├── metadata_val.jsonl       # Validation split metadata
├── metadata_test.jsonl      # Test split metadata
├── metadata.jsonl           # Full metadata
├── cic_downstream.jsonl     # culture-informed image captioning task
├── cvqa_downstream.jsonl    # culture-centric visual question answering task
└── wiki_documents.jsonl     # Corpus of Wikipedia articles for retrieval
```

## Schema

### Metadata (`metadata_*.jsonl`)

Each line is a JSON object representing a data sample:

- `file_name`: Path to the image file (e.g., `./ravenea/images/ccub_101_China_38.jpg`).
- `country`: Country of origin for the cultural content.
- `task_type`: Task category (e.g., `cIC` for image captioning/QA).
- `category`: Broad cultural category (e.g., `Daily Life`).
- `human_captions`: Human-written caption describing the image.
- `questions`: List of questions associated with the image.
- `options`: Multiple-choice options for the questions.
- `answers`: Correct answers for the questions.
- `enwiki_ids`: List of relevant Wikipedia article IDs.
- `culture_relevance`: Score or indicator of cultural relevance.

### Wikipedia Corpus (`wiki_documents.jsonl`)

Contains the knowledge base for retrieval:

- `id`: Unique identifier for the article (e.g., `enwiki/65457597`).
- `text`: Full text content of the Wikipedia article.
- `date_modified`: Last modification date of the article.

## Usage

### Download the Dataset

Please download the dataset then unzip it to the current directory.
```python
from huggingface_hub import hf_hub_download

local_path = hf_hub_download(
    repo_id="jaagli/ravenea",
    filename="./ravenea.zip",
    repo_type="dataset",
    local_dir="./",
)
print(f"File downloaded to: {local_path}")
```

### Loading the Data
You can load the dataset using standard Python libraries.:

```python
import json
from pathlib import Path

def load_jsonl(file_path):
    data = []
    with open(file_path, 'r', encoding='utf-8') as f:
        for line in f:
            data.append(json.loads(line))
    return data

# Load metadata
train_data = load_jsonl("./ravenea/metadata_train.jsonl")
test_data = load_jsonl("./ravenea/metadata_test.jsonl")

# Load Wikipedia corpus
wiki_docs = load_jsonl("./ravenea/wiki_documents.jsonl")
doc_id_to_text = {doc['id']: doc['text'] for doc in wiki_docs}

# Example: Accessing a sample
sample = train_data[0]
print(f"Image: {sample['file_name']}")
print(f"Caption: {sample['human_captions']}")
print(f"Docs: {sample['enwiki_ids']}")
```

## BibTeX Citation
```bibtex
@inproceedings{
li2026ravenea,
title={{RAVENEA}: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding},
author={Jiaang Li and Yifei Yuan and Wenyan Li and Mohammad Aliannejadi and Daniel Hershcovich and Anders S{\o}gaard and Ivan Vuli{\'c} and Wenxuan Zhang and Paul Pu Liang and Yang Deng and Serge Belongie},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=4zAbkxQ23i}
}
```