clips/SynEmbedNL

---
dataset_info:
features:
- name: task_type
dtype: string
- name: task_desc
dtype: string
- name: model
dtype: string
- name: query
dtype: string
- name: pos
dtype: string
- name: neg
dtype: string
- name: pos_scores
list: float64
- name: neg_scores
list: float64
- name: query-id
dtype: string
- name: pos-id
dtype: string
- name: neg-id
dtype: string
splits:
- name: train
num_bytes: 771179285
num_examples: 499252
download_size: 411040575
dataset_size: 771179285
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- feature-extraction
language:
- nl
tags:
- synthetic
size_categories:
- 100K<n<1M
license: mit
---

### Synthetic Dataset for Training Dutch Text Embedding Models
This is the synthetic dataset used for training Dutch embedding models as described in [MTEB-NL and E5-NL: Embedding Benchmark and Models for Dutch](https://arxiv.org/abs/2509.12340).
Each sample contains the following fields:
- **task_type**: Type of the embedding task; one of the five categories:
- sl (short-long): retrieval
- ls (long-short): classification
- ss (short-short): clustering
- ll (long-long): clustering
- sts (semantic text similarity): semantic text similarity
- **task_desc**: The general prompt used for describing the task
- **model**: The LLM used for generation
- **query**: The generated query
- **pos**: The generated *positive* document
- **neg**: The generated *negative* document
- **pos_scores**: The relevance score of the positive document, as calculated by Qwen3-Reranker
- **neg_scores**: The relevance score of the negative document, as calculated by Qwen3-Reranker

### Trained Models
Trained models can be accessed [here](https://huggingface.co/collections/clips/e5-nl).

## Citation Information

If you find our paper, benchmark or models helpful, please consider cite as follows:
```latex
@misc{banar2025mtebnle5nlembeddingbenchmark,
title={MTEB-NL and E5-NL: Embedding Benchmark and Models for Dutch},
author={Nikolay Banar and Ehsan Lotfi and Jens Van Nooten and Cristina Arhiliuc and Marija Kliocaite and Walter Daelemans},
year={2025},
eprint={2509.12340},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2509.12340},
}
```
[//]: # (https://arxiv.org/abs/2509.12340)