impresso-project / histlux-gte-multilingual-base

Luxembourgish adaptation of Alibaba-NLP/gte-multilingual-base

This is a sentence-transformers model finetuned from Alibaba-NLP/gte-multilingual-base further adapted to support Historical and Contemporary Luxembourgish. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for (cross-lingual) semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

This model is specialised to perform cross-lingual semantic search to and from Historical/Contemporary Luxembourgish. This model would be particularly useful for libraries and archives that want to perform semantic search and longitudinal studies within their collections.

This is an Alibaba-NLP/gte-multilingual-base model that was further adapted by (Michail et al., 2025)

Limitations

Whilst this model natively supports inputs up to 8192, all of our evaluations are on sentence level so there are no guarantees on it's longer text embedding capabilities of Historical Luxembourgish.

We also release a model that performs better (18pp) on ParaLUX. If finding monolingual exact matches within adversarial collections is of at-most importance, please use histlux-paraphrase-multilingual-mpnet-base-v2

Model Description

• Model Type: GTE-Multilingual-Base
• Base model: Alibaba-NLP/gte-multilingual-base
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • LB-EN (Hist-TR, RTL-M)

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('impresso-project/histlux-gte-multilingual-base', trust_remote_code=True)
embeddings = model.encode(sentences)
print(embeddings)

Evaluation Results

Metrics

(see introducing paper) Historical Bitext Mining (Accuracy):

LB -> FR: 96.8

FR -> LB: 96.9

LB -> EN: 97.2

EN -> LB: 97.2

LB -> DE: 98.0

DE -> LB: 91.8

Contemporary LB (Accuracy): ParaLUX: 62.82

SIB-200(LB): 62.16

Training Details

Training Dataset

The parallel sentences data mix is the following:

impresso-project/HistLuxAlign:

• LB-FR (x20,000)
• LB-EN (x20,000)
• LB-DE (x20,000)

fredxlpy/LuxAlign:

• LB-FR (x40,000)
• LB-EN (x20,000)

Total: 120 000 Sentence pairs in mixed batches of size 8

Contrastive Training

The model was trained with the parameters:

**Loss**:

`sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss` with parameters:

{'scale': 20.0, 'similarity_fct': 'cos_sim'}

Parameters of the fit()-Method:

{ "epochs": 1, "evaluation_steps": 520, "max_grad_norm": 1, "optimizer_class": "<class 'torch.optim.adamw.AdamW'>", "optimizer_params": { "lr": 2e-05 }, "scheduler": "WarmupLinear", }

Citation

BibTeX

Adapting Multilingual Embedding Models to Historical Luxembourgish (introducing paper)

@misc{michail2025adaptingmultilingualembeddingmodels,
      title={Adapting Multilingual Embedding Models to Historical Luxembourgish}, 
      author={Andrianos Michail and Corina Julia Raclé and Juri Opitz and Simon Clematide},
      year={2025},
      eprint={2502.07938},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.07938}, 
}

Original Multilingual GTE Model

@inproceedings{zhang2024mgte,
  title={mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval},
  author={Zhang, Xin and Zhang, Yanzhao and Long, Dingkun and Xie, Wen and Dai, Ziqi and Tang, Jialong and Lin, Huan and Yang, Baosong and Xie, Pengjun and Huang, Fei and others},
  booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  pages={1393--1412},
  year={2024}
}