e5-sk-large

e5-sk-large is a Slovak text embedding model (365M parameters, 1024-dimensional embeddings) built by applying vocabulary trimming and fine-tuning to multilingual-e5-large. It achieves competitive performance with proprietary embedding APIs on SkMTEB — the first comprehensive Slovak text embedding benchmark — while being 35% smaller than the original model and fully locally deployable.

Released as part of the SkMTEB project (paper · GitHub · collection).

For a smaller, faster variant, see e5-sk-small (45M parameters).

---

Model Details

Usage

This model follows the standard E5 prefix convention: prepend query: to queries and passage: to documents during retrieval. For symmetric tasks (STS, clustering, classification), no prefix is needed.

With sentence-transformers

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer
 
model = SentenceTransformer("slovak-nlp/e5-sk-large")
 
# Retrieval
query_embedding = model.encode("query: Čo je hlavné mesto Slovenska?")
passage_embedding = model.encode("passage: Bratislava je hlavné a najväčšie mesto Slovenska.")
similarity = model.similarity(query_embedding, passage_embedding)
print(similarity)  # tensor([[0.9269]])
 
# Batch encoding
sentences = [
    "query: Aké je počasie v Bratislave?",
    "passage: V Bratislave je dnes slnečno a teplo.",
    "passage: Bratislava leží na brehu Dunaja.",
]
embeddings = model.encode(sentences)
print(embeddings.shape)  # (3, 1024)

With transformers directly

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel
 
def average_pool(last_hidden_states, attention_mask):
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
 
tokenizer = AutoTokenizer.from_pretrained("slovak-nlp/e5-sk-large")
model = AutoModel.from_pretrained("slovak-nlp/e5-sk-large")
 
texts = [
    "query: Čo je hlavné mesto Slovenska?",
    "passage: Bratislava je hlavné a najväčšie mesto Slovenska.",
]
 
batch_dict = tokenizer(texts, max_length=512, padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
    outputs = model(**batch_dict)
 
embeddings = average_pool(outputs.last_hidden_state, batch_dict["attention_mask"])
embeddings = F.normalize(embeddings, p=2, dim=1)
print((embeddings[0] @ embeddings[1]).item())

Prefix guide

---

Training

How it was built

Step 1 — Vocabulary Trimming. Before fine-tuning, Vocabulary Trimming (Ushio et al., 2023) was applied to multilingual-e5-large to remove tokens irrelevant to Slovak. Token frequencies were computed on FineWeb2-Slovak, a quality-filtered Slovak web corpus, and the top 60K tokens (out of 250K) were retained. This reduced the model from 560M to 365M parameters (35% reduction) without meaningful performance loss.

Step 2 — Fine-tuning. The trimmed model was fine-tuned on curated Slovak datasets from the skLEP benchmark:

Training configuration

---

Evaluation: SkMTEB Results

Evaluated on SkMTEB — 31 datasets across 7 task types. Scores are percentages (higher is better).

e5-sk-large is practically equivalent to text-embedding-3-large (TOST equivalence test: 90% CI within ±2 points), while being open-weight, locally deployable, and free to run. Cross-lingual Slovak–English and Slovak–Czech bitext mining performance is preserved within 1 F1 point compared to the original multilingual-e5-large.

---

Intended Uses

• Semantic search and retrieval-augmented generation (RAG) over Slovak text
• Semantic textual similarity (STS)
• Text clustering and classification via embedding features
• Cross-lingual retrieval (Slovak–English, Slovak–Czech)
• Local deployment where API latency or cost is a concern

Limitations

• Optimised for Slovak; cross-lingual transfer to non-Slavic languages is not evaluated.
• Vocabulary trimming removes non-Slovak tokens; performance on heavily code-mixed text may be reduced.
• Training data skews toward news, parliamentary, and encyclopedic domains.
• Max sequence length during fine-tuning is 256 tokens (underlying architecture supports up to 512).

---

Citation

@inproceedings{suppa2025skmteb,
  title     = {{SkMTEB}: {Slovak} Massive Text Embedding Benchmark and Model Adaptation},
  author    = {{\v{S}}uppa, Marek and Ridzik, Andrej and Hl{\'a}dek, Daniel and
               Kna{\v{z}}ekov{\'a}, Nat{\'a}lia and Ondrejov{\'a}, Vikt{\'o}ria},
  year      = {2025},
  eprint    = {2606.13647},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2606.13647}
}
 
@inproceedings{reimers-2019-sentence-bert,
  title     = {Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks},
  author    = {Reimers, Nils and Gurevych, Iryna},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
  year      = {2019},
  publisher = {Association for Computational Linguistics},
  url       = {https://arxiv.org/abs/1908.10084}
}

---

License

MIT