SkMTEB
Slovak NLP · Text Embeddings ACL 2026

SkMTEB: a Massive Text Embedding Benchmark for Slovak

The first comprehensive embedding benchmark for Slovak — and two open models that run locally yet match proprietary APIs.

Marek Šuppaα,β   Andrej Ridzikδ   Daniel Hládekγ   Natália Kňažekováα,δ   Viktória Ondrejováβ
αComenius University in Bratislava  ·  βCisco Systems  ·  γTechnical University of Košice  ·  δKempelen Institute of Intelligent Technologies

A 45M-parameter Slovak model reaches 91% of the best 560M multilingual model — and runs right here in your browser

31
datasets across 7 task types
~4×
the Slovak depth of MMTEB (8 tasks)
31
embedding models evaluated
−62%
size cut with no quality loss
The benchmark

What's inside SkMTEB

Slovak — a West Slavic language with ~5 million speakers — had no embedding benchmark. MMTEB covers it with just 8 tasks, mostly multilingual subsets. SkMTEB collates 31 datasets spanning news, government, medical, social-media, and encyclopedic Slovak (2000–2025), including 7 datasets created specifically for this work. Only 6 overlap with MMTEB, so the two are complementary.

5
Retrieval
SK-QuAD, news summaries-as-retrieval, FAQ
3
Reranking
Pharmacist Q&A, SK-QuAD hard negatives
7
Classification
Sentiment, hate speech, parliament, topic
5
Clustering
News by tag / URL / category, SIB-200
6
Bitext Mining
SK↔EN, SK↔CS: Flores, NTREX, Tatoeba
3
Pair Classification
NLI, RTE, Demagog fact-checking
2
STS
SlovakSTS, synthetic SlovakSumSTS
The models

Open Slovak embeddings, built efficiently

We adapt Multilingual E5 by vocabulary trimming (250K → 60K tokens, kept by Slovak frequency) followed by fine-tuning on curated skLEP data. Training takes under one GPU-hour per model on a single H100. Both are open-weight and locally deployable — no API, no data leaving the machine.

e5-sk-smallruns in-browser
Params
45M
vs. mE5-small
−62%
SkMTEB avg
70.56

On par with text-embedding-3-small (70.48) at a fraction of the footprint. Small enough to run directly in a web browser.

🤗 slovak-nlp/e5-sk-small →
e5-sk-largeAPI-grade, local
Params
365M
vs. mE5-large
−35%
SkMTEB avg
74.70

Comparable to text-embedding-3-large (75.07); TOST equivalence within ±2 points. An open model you can host yourself for Slovak semantic search and RAG.

🤗 slovak-nlp/e5-sk-large →
Try it live

Run e5-sk-small in your browser

No server, no API key. Load the 45M-parameter model once (~45 MB, then cached) and it runs entirely on your device. Ask a question in Slovak and it ranks the snippets by meaning, even when none of the words match.

Downloaded once and cached · nothing leaves your device · powered by transformers.js

The signature result

Quality per parameter

Average SkMTEB score vs. parameter count (log scale). The Pareto frontier connects models that no smaller model beats. Both Slovak models sit on or near the frontier — and beyond ~600M, scale stops paying off for Slovak.

Our Slovak models Pareto-optimal Other open models Pareto frontier
Full results

Leaderboard

All 31 models on SkMTEB (percent). All = mean over 31 datasets; Type = unweighted mean over task types. Click any header to sort; best per column is highlighted. Multilingual E5-large-instruct leads overall, with Gemini close behind.

Tip: sort by a task column to find the per-task leader.
SkMTEB results. Btxt = Bitext Mining · Clf = Classification · Clust = Clustering · PrClf = Pair Classification · Rrnk = Reranking · Rtrvl = Retrieval · STS = Semantic Textual Similarity.
Why it works

Trimming is (almost) free

Vocabulary trimming removes tokens that Slovak never uses. It cuts size by 35–62% while quality holds steady or improves slightly, and transfer to English and Czech stays within 1 F1 point.

VariantVTFTSizeAvgΔ
mE5-small118M70.32
+ VT45M70.45+0.13
+ VT + FT45M70.56+0.24
+ VT + FT + prompt45M71.07+0.75
mE5-large560M74.25
+ VT365M74.56+0.31
+ VT + FT365M74.70+0.45
+ VT + FT + prompt365M74.72+0.47

Ablation: vocabulary trimming (VT), fine-tuning (FT), and query:/passage: prompts. Δ is vs. the untrimmed baseline.

Bitext pairE5-sm origtrimE5-lg origtrim
Flores ces–slk99.8799.7499.8799.87
Flores eng–slk95.5895.72100.0100.0
NTREX ces–slk98.0898.1199.2299.27
NTREX eng–slk94.6495.5698.8098.66
Tatoeba slk–eng82.6982.5693.2293.22
Opus slk–eng77.0876.9586.2186.27

Cross-lingual transfer is preserved: mean change 0.25 F1 (small) and 0.04 F1 (large) after trimming.

Use it

Run the benchmark

SkMTEB ships through the standard MTEB library. Evaluate any model in a few lines.

# pip install mteb
import mteb

model = mteb.get_model("slovak-nlp/e5-sk-small")
tasks = mteb.get_benchmark("SkMTEB")
results = mteb.MTEB(tasks=tasks).run(model, output_folder="results/")
Cite

BibTeX

@inproceedings{suppa-etal-2026-skmteb,
  title     = {{SkMTEB}: {S}lovak Massive Text Embedding Benchmark and Model Adaptation},
  author    = {{\v{S}}uppa, Marek and Ridzik, Andrej and Hl{\'a}dek, Daniel and
               Kn{\v{a}}{\v{z}}ekov{\'a}, Nat{\'a}lia and Ondrejov{\'a}, Vikt{\'o}ria},
  booktitle = {Proceedings of ACL},
  year      = {2026},
  eprint    = {2606.13647},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2606.13647},
}