README.md

CC3M-eu (Basque Translation)

📚 Overview

CC3M-eu is a Basque version of the Conceptual Captions 3M dataset. It consists of approximately 3.3 million image-description pairs where the original English captions have been translated into Basque using the mt-hitz-en/eu specialized translation pipeline.

Important: This is not the official dataset. It is an independent community translation effort designed to facilitate the training of Vision-Language Models (VLMs) and CLIP-style encoders for the Basque language.

✍️ Authors & Acknowledgements

• Original dataset: Conceptual Captions — © Google LLC
• Basque translation & curation: <Lukas Arana / HiTZ>, 2025
• Translation Engine: Neural Machine Translation via mt-hitz-en/eu (HiTZ Center).

If you use this Basque split, please cite both the original Conceptual Captions paper and this translation work.

📁 Dataset Schema

The schema has been adapted to include both the source and the translated content:

1. id: Unique identifier for each sample.
2. url: The original source URL for the image.
3. caption_en: The original English description.
4. caption_eu: The generated Basque translation.

🔧 How We Built It

1. Extraction: English captions were pulled from the official CC3M TSV files.
2. Translation: Each caption was translated using the mt-hitz-en/eu model, which is specifically optimized for English-to-Basque scientific and general domain text.
3. Cleanup: Applied basic post-processing to handle HTML entities and formatting artifacts.

Note: No images were hosted or modified; only the textual metadata is provided.

🚦 Limitations & Ethical Considerations

• Non-official: This version has not been audited by Google; semantic drift may occur during translation.
• Link Stability: Like the original CC3M, many image URLs may be dead or lead to 404 errors.
• Biases: The dataset may inherit or amplify social biases present in the original English data or the NMT model.

💻 Quick Start

from datasets import load_dataset

# Load the Basque CC3M dataset
ds = load_dataset("lukasArana/CC3M-eu", split="train")

# Accessing a sample
print(ds[0]["caption_eu"])