wikipedia-40-langs-with-embeddings

A balanced multilingual Wikipedia embedding dataset covering 40 languages, released as part of the reproducibility artifacts for the paper:

Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation
M. Ali Bayram, Banu Diri, Savaş Yıldırım
arXiv: https://arxiv.org/abs/2605.29992

This dataset contains Wikipedia text samples together with precomputed teacher embeddings generated using google/embeddinggemma-300m. It was created to support offline embedding distillation experiments for adapting multilingual sentence embedding models, especially for Turkish and other lower-resource or morphologically rich languages.

The dataset was used in the development of embeddingmagibu-200m, a Turkish-focused sentence embedding model trained through cross-lingual tokenizer surgery, model cloning, and offline embedding distillation.

Dataset Description

This dataset contains 580,000 Wikipedia text samples from a balanced 40-language corpus. Each sample includes language metadata, article title, text content, source URL, and two teacher embedding representations extracted from the teacher model.

The main purpose of the dataset is to enable efficient student-model training without requiring online teacher inference during distillation. By storing precomputed teacher vectors, the student model can be trained directly against embedding targets using cosine similarity or related embedding-space objectives.

Dataset Features

Each row contains the following fields:

Field	Type	Description
lang	string	Language code of the Wikipedia sample
title	string	Wikipedia article title
text	string	Text passage used for embedding generation
url	string	Source Wikipedia URL
teacher_embedding_final	list[float64]	Final teacher embedding vector
teacher_embedding_pre_dense	list[float64]	Teacher representation before the dense projection layer

Dataset Size

Split	Examples
train	580,000

Approximate dataset size:

Metric	Value
Dataset size	10.97 GB
Download size	7.62 GB

Intended Use

This dataset is intended for:

• Offline embedding distillation
• Multilingual sentence embedding research
• Turkish-focused embedding model adaptation
• Cross-lingual tokenizer and vocabulary adaptation experiments
• Low-resource and morphologically rich language NLP research
• Semantic search and retrieval model development
• Reproducibility of the embeddingmagibu-200m training pipeline

Relation to the Paper

This dataset is one of the released artifacts associated with the following paper:

Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation
arXiv: https://arxiv.org/abs/2605.29992

In the paper, the dataset is used for the offline distillation stage. Instead of running the teacher model during student training, teacher embeddings are precomputed and stored in this dataset. This makes training faster, cheaper, and more reproducible.

The broader pipeline consists of:

1. Constructing a Turkish-optimized multilingual tokenizer
2. Cloning the teacher embedding model with a compatible embedding table
3. Distilling from precomputed teacher embeddings using this dataset

Usage

You can load the dataset with the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("alibayram/wikipedia-40-langs-with-embeddings")

print(dataset)
print(dataset["train"][0])

Example fields:

sample = dataset["train"][0]

text = sample["text"]
lang = sample["lang"]
final_embedding = sample["teacher_embedding_final"]
pre_dense_embedding = sample["teacher_embedding_pre_dense"]

Example Use Case: Offline Distillation

A student embedding model can be trained by encoding the text field and minimizing the distance between the student embedding and the stored teacher embedding.

For example, a cosine-similarity objective can be used between:

• Student model output embedding
• teacher_embedding_final

This avoids repeated teacher-model forward passes and enables efficient reproduction of the distillation setup described in the paper.

Associated Model and Resources

This dataset is associated with the embeddingmagibu-200m model and the open-source training tools released for the paper.

Paper:

• https://arxiv.org/abs/2605.29992

Project page:

• https://huggingface.co/spaces/magibu/embeddingmagibu-200m

GitHub repository:

• https://github.com/malibayram/embedding-trainer

Limitations

This dataset contains Wikipedia-derived text and teacher-model embeddings. Therefore:

• The quality of the embeddings depends on the teacher model used to generate them.
• The dataset reflects the linguistic and topical distribution of the selected Wikipedia samples.
• It should not be interpreted as a general-purpose benchmark dataset.
• It is primarily designed as a training and reproducibility artifact for offline embedding distillation.

Citation

If you use this dataset, please cite the associated paper:

@article{bayram2026embeddingmagibu,
  title={Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation},
  author={Bayram, M. Ali and Diri, Banu and Yıldırım, Savaş},
  journal={arXiv preprint arXiv:2605.29992},
  year={2026},
  url={https://arxiv.org/abs/2605.29992}
}