gorkemergune/stsb-tr

Turkish STS — sentence pairs scored with magibu/embeddingmagibu-200m

A Turkish Semantic Textual Similarity (STS) dataset: each row is a pair of sentences plus a similarity score. Scores come from the magibu/embeddingmagibu-200m sentence-embedding model, computed as cosine similarity over L2-normalized embeddings — the exact method used by the reference Space.

The set deliberately spans the full similarity range, with many near-zero (unrelated) pairs, so it can be used to evaluate or calibrate similarity thresholds — not just high-similarity paraphrase detection.

---

At a glance

Total pairs	1037
Manually collected (manual)	38
Synthetic (synthetic)	999
Splits	train 933 / test 104 (90 / 10)
Language	Turkish (tr)
Score	cosine similarity, ≈ -0.04 … 0.97

Score distribution

Score band	Pairs
0.0 – 0.2 (unrelated)	413
0.2 – 0.4	197
0.4 – 0.6	51
0.6 – 0.8	138
0.8 – 1.0 (near-identical)	238

By pair type (synthetic):

pair_type	n	mean	min	max
unrelated	450	0.148	−0.039	0.456
related	250	0.406	0.068	0.938
paraphrase	299	0.844	0.446	0.969

---

Columns

Column	Description
sentence1, sentence2	The two compared Turkish sentences
score	Cosine similarity from magibu/embeddingmagibu-200m
source	manual (hand-collected) or synthetic
pair_type	manual / unrelated / related / paraphrase
topic	Topic of the synthetic pair (for unrelated, both topics as a|b)
split	train or test

Usage

from datasets import load_dataset

ds = load_dataset("gorkemergune/stsb-tr")
print(ds)
print(ds["train"][0])

# e.g. keep only strongly-similar pairs
paraphrases = ds["train"].filter(lambda r: r["score"] >= 0.8)

---

How it was built

Manual pairs (38). Real Turkish sentence pairs (news headlines and their reworded versions) collected by hand and scored with the model.

Synthetic pairs (999). Template-generated sentences in the style of Turkish news pages, across nine topics: magazine/celebrity, sports, economy, politics, weather, crime & accidents, health, technology, world. Pairs are built at three relatedness levels so scores span the whole range:

• unrelated — two sentences from different topics → near-zero score
• related — same topic, different event → low/medium score
• paraphrase — the same event phrased two ways → high score

Every pair — manual and synthetic alike — is scored by the same model, so the column is internally consistent. Exact duplicates and identical-sentence pairs were removed.

Limitations

• The synthetic sentences are not real news content; they imitate the style of the referenced outlets and were produced from templates. No real article text is reproduced.
• score is a model output, not a human judgment. It reflects magibu/embeddingmagibu-200m's notion of similarity and inherits its biases. Treat it as a silver label, not gold.
• Synthetic paraphrases are cleaner and more regular than real-world text, so the paraphrase band may be easier than natural data.

License

Released under the GNU General Public License v3.0 (GPLv3). If you use it, please also credit the underlying model magibu/embeddingmagibu-200m.