STASIS — Semantic Textual Similarity datasets, unified

Modernised release of the alvations/stasis collection of Semantic Textual Similarity (STS) shared-task data, repackaged into parquet under one unified schema so every subset can be loaded the same way.

Replaces the original graphlab-based wrappers (graphlab is defunct) with stdlib + pyarrow only.

Configs

Config	Source	Rows	Splits	Score scale
sts2012	SemEval-2012 STS	5,276	train (2,212), test (3,064)	0–5
sts2013	SemEval-2013 STS	1,505	test	0–5
sts2014	SemEval-2014 STS	3,747	test	0–5
sts2015	SemEval-2015 STS	2,984	test	0–5
sts2016	SemEval-2016 STS	1,169	test (gold-scored subset)	0–5
sts2017	SemEval-2017 STS (multilingual)	2,000	test	— (scores not released into this repo)
sick	SICK relatedness	9,927	train (4,500), validation (500), test (4,927)	1–5
clss-paragraph2sentence	SemEval-2014 CLSS	1,034	train (500), validation (34), test (500)	0–4
clss-sentence2phrase	SemEval-2014 CLSS	1,036	train (500), validation (36), test (500)	0–4
clss-phrase2word	SemEval-2014 CLSS	1,038	train (500), validation (38), test (500)	0–4
clss-word2sense	SemEval-2014 CLSS	1,048	train (500), validation (48), test (500)	0–4
all-sts-en	concat of sts2012–sts2016	14,681	train	0–5

Usage

from datasets import load_dataset

# any STS year — same call shape
ds = load_dataset("alvations/stasis", "sts2014", split="test")
print(ds[0])
# {'sent1': '...', 'sent2': '...', 'score': 4.4, 'year': 2014, ...}

# all English STS in one go
ds_all = load_dataset("alvations/stasis", "all-sts-en", split="train")
print(len(ds_all))   # 14681

# SICK with entailment labels
sick = load_dataset("alvations/stasis", "sick", split="test")
print(sick[0]["entailment"])    # 'NEUTRAL' / 'ENTAILMENT' / 'CONTRADICTION'

# multilingual STS 2017 (Arabic, Spanish, English, Turkish pairs)
sts17 = load_dataset("alvations/stasis", "sts2017", split="test")
print(sts17[0]["lang1"], sts17[0]["lang2"])    # e.g. 'ar' 'en'

Unified schema

All configs share one schema — makes cross-subset comparison trivial:

Column	Type	Description
sent1	string	First segment (sentence, phrase, or word — depends on subset).
sent2	string	Second segment.
score	float32	Gold similarity score. Scale depends on subset (STS 0–5, SICK 1–5, CLSS 0–4). Null where unreleased (e.g. STS2017 in this repo).
year	int32	Dataset year (2012–2017). SICK = 2014, CLSS = 2014.
subset	string	Sub-source within the year (MSRpar, headlines, deft-news, track1.ar-ar, paragraph2sentence, …).
split	string	"train" / "validation" / "test".
lang1	string	ISO-2 language code of sent1. Always en except STS2017.
lang2	string	ISO-2 language code of sent2.
pair_id	string	Upstream pair identifier (SICK pair_ID; CLSS pair id). Null elsewhere.
entailment	string	SICK only: NEUTRAL / ENTAILMENT / CONTRADICTION. Null elsewhere.

Notes on individual configs

• STS 2012: includes the official train split (MSRpar, MSRvid, SMTeuroparl) plus the gold test set with the surprise domains (OnWN, SMTnews).
• STS 2013–2015: test/gold splits only — those years didn't release new training data.
• STS 2016: the upstream STS2016 input file has 9,183 sentence pairs but only 1,169 carry gold scores (the rest were unannotated test material). The unannotated rows are dropped here.
• STS 2017: multilingual tracks (ar-ar, ar-en, es-es, es-en ×2, en-en, tr-en). Each row stores the original-language sentences in sent1/sent2; English translations from the upstream sts2017.csv are not surfaced as a separate column (open issue if you'd like them).
• SICK: standard 4,500 / 500 / 4,927 split; entailment column carries the entailment judgment alongside the relatedness score.
• CLSS: 4 cross-lexical-level configs from SemEval-2014 Task 3. For clss-word2sense, sent1 is word#pos and sent2 is the WordNet sense's human-readable form (the raw sense_id is in the upstream input but elided here — file a request if you need it).

Modernisation notes (vs upstream)

This release rebuilds the data with:

• pure stdlib + pyarrow (no graphlab / turicreate — both defunct);
• single unified parquet schema across every subset;
• explicit lang1/lang2/split/subset columns instead of relying on tabular conventions;
• empty-score rows from the upstream sts.csv (STS2016 unannotated subset) dropped, with the count tracked above;
• one HF config per source year + a default-style all-sts-en aggregate.

The source files were not modified — the parquet shards are a faithful flatten of the upstream sts.csv and per-subset folders.

Citations

Cite the original tasks alongside this repackaging. The most important ones:

@inproceedings{agirre-etal-2012-semeval,
  title = {{S}em{E}val-2012 Task 6: A Pilot on Semantic Textual Similarity},
  author = {Agirre, Eneko and Cer, Daniel and Diab, Mona and Gonzalez-Agirre, Aitor},
  booktitle = {SemEval-2012},
  year = {2012},
}

@inproceedings{agirre-etal-2013-sem,
  title = {*{SEM} 2013 shared task: Semantic Textual Similarity},
  author = {Agirre, Eneko and Cer, Daniel and Diab, Mona and Gonzalez-Agirre, Aitor and Guo, Weiwei},
  booktitle = {*SEM 2013},
  year = {2013},
}

@inproceedings{agirre-etal-2014-semeval,
  title = {{S}em{E}val-2014 Task 10: Multilingual Semantic Textual Similarity},
  author = {Agirre, Eneko and Banea, Carmen and Cardie, Claire and Cer, Daniel and Diab, Mona and Gonzalez-Agirre, Aitor and Guo, Weiwei and Mihalcea, Rada and Rigau, German and Wiebe, Janyce},
  booktitle = {SemEval-2014},
  year = {2014},
}

@inproceedings{agirre-etal-2015-semeval,
  title = {{S}em{E}val-2015 Task 2: Semantic Textual Similarity, {E}nglish, {S}panish and Pilot on Interpretability},
  author = {Agirre, Eneko and others},
  booktitle = {SemEval-2015},
  year = {2015},
}

@inproceedings{agirre-etal-2016-semeval,
  title = {{S}em{E}val-2016 Task 1: Semantic Textual Similarity, Monolingual and Cross-Lingual Evaluation},
  author = {Agirre, Eneko and Banea, Carmen and Cer, Daniel and Diab, Mona and Gonzalez-Agirre, Aitor and Mihalcea, Rada and Rigau, German and Wiebe, Janyce},
  booktitle = {SemEval-2016},
  year = {2016},
}

@inproceedings{cer-etal-2017-semeval,
  title = {{S}em{E}val-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual Focused Evaluation},
  author = {Cer, Daniel and Diab, Mona and Agirre, Eneko and Lopez-Gazpio, I{\~n}igo and Specia, Lucia},
  booktitle = {SemEval-2017},
  year = {2017},
}

@inproceedings{marelli-etal-2014-sick,
  title = {A {SICK} cure for the evaluation of compositional distributional semantic models},
  author = {Marelli, Marco and Menini, Stefano and Baroni, Marco and Bentivogli, Luisa and Bernardi, Raffaella and Zamparelli, Roberto},
  booktitle = {LREC 2014},
  year = {2014},
}

@inproceedings{jurgens-etal-2014-semeval,
  title = {{S}em{E}val-2014 Task 3: Cross-Level Semantic Similarity},
  author = {Jurgens, David and Pilehvar, Mohammad Taher and Navigli, Roberto},
  booktitle = {SemEval-2014},
  year = {2014},
}

If you want to cite the wrapper itself, the STASIS repo suggests citing:

@inproceedings{han-etal-2015-samsung,
  title = {Samsung: Align-and-Differentiate Approach to Semantic Textual Similarity},
  author = {Han, Lushan and Kashyap, Abhay L. and Finin, Tim and Mayfield, James and Weese, Jonathan and others},
  booktitle = {*SEM 2015},
  url = {http://www.aclweb.org/anthology/S15-2015},
  year = {2015},
}

License

Each subset inherits its upstream license:

• STS 2012–2017: research use as released by the SemEval organisers.
• SICK: CC BY-NC-SA 3.0.
• CLSS: research use as released by the SemEval-2014 Task 3 organisers.

This repackaging adds no additional restrictions but downstream users must respect the per-subset terms.