Dataset Viewer
Auto-converted to Parquet Duplicate
page_id
int64
title
string
title_en
string
topic
string
text
large_string
formulas
int32
num_chunks
int32
translated_chunks
int32
edge_chunks_removed
int32
interior_gaps
int32
llama32_tokens
int64
source_url
string
source_revision
int64
source_revision_url
string
source_date_modified
string
translator
string
pipeline
string
license
string
39
Albedo
Albedo
Earth and environment
Albedo Albedo (/ælˈbiːdoʊ/ al-BEE-doh; Latince albedo 'beyazlık' kelimesinden) bir cismin difüz olarak yansıttığı güneş ışığının kesridir. 0'dan (tüm gelen radyasyonu soğan siyah bir cisme karşılık gelen) 1'e (tüm gelen radyasyonu yansıtan bir cisme karşılık gelen) kadar bir ölçekte ölçülür. Yüzey albedosu, bir yüzeyi...
38
8
8
0
0
7,442
https://en.wikipedia.org/wiki/Albedo
1,278,647,862
https://en.wikipedia.org/w/index.php?oldid=1278647862
2025-03-03T19:03:58Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
572
Tarım bilimi
Agricultural science
Earth and environment
Tarım bilimi Tarım bilimi (kısa adıyla agriscience), tarımın pratiğinde ve anlaşılmasında kullanılan kesin, doğal, ekonomik ve sosyal bilimlerin kısımlarını kapsayan geniş bir disiplinler arası biyoloji alanıdır. Tarım bilimi uzmanlarına tarım bilimcileri veya ziraatçılar denir. Tarih 18. yüzyılda Johann Friedrich M...
0
2
2
0
0
1,380
https://en.wikipedia.org/wiki/Agricultural_science
1,277,462,037
https://en.wikipedia.org/w/index.php?oldid=1277462037
2025-02-24T19:55:07Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
580
Astronom
Astronomer
Space
Astronom Bir astronom, astronomi alanında, Dünya'nın kapsamı dışında belirli bir soruya veya alana odaklanan bir bilim insanıdır. Astronomlar, yıldızlar, gezegenler, uydular, kuyruklu yıldızlar ve galaksiler gibi astronomik nesneleri, gözlemsel (veri analiz ederek) veya teorik astronomi yoluyla gözlemler. Astronomları...
0
2
2
0
0
1,421
https://en.wikipedia.org/wiki/Astronomer
1,272,660,334
https://en.wikipedia.org/w/index.php?oldid=1272660334
2025-01-29T19:06:28Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
586
ASCII
ASCII
Computing
ASCII ASCII (/ˈæskiː/ ASS-kee), American Standard Code for Information Interchange (Bilgi Değişimi için Amerikan Standart Kodu) ifadesinin kısaltmasıdır ve elektronik iletişim için bir karakter kodlama standardıdır. ASCII kodları, bilgisayarlarda, telekomünikasyon ekipmanlarında ve diğer cihazlarda metni temsil eder. ...
0
11
10
0
1
9,083
https://en.wikipedia.org/wiki/ASCII
1,279,728,946
https://en.wikipedia.org/w/index.php?oldid=1279728946
2025-03-10T05:59:02Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
612
Aritmetik ortalama
Arithmetic mean
Mathematics
Aritmetik ortalama Matematik ve istatistikte, aritmetik ortalama ( /ˌærɪθˈmɛtɪk/ arr-ith-MET-ik), aritmetik ortalama veya bağlam açık olduğunda sadece ortalama, bir sayı koleksiyonunun toplamının, koleksiyondaki sayıların sayısına bölünmesiyle elde edilen değerdir. Koleksiyon genellikle bir deneyden, gözlemsel bir çal...
74
4
4
0
0
3,015
https://en.wikipedia.org/wiki/Arithmetic_mean
1,278,223,585
https://en.wikipedia.org/w/index.php?oldid=1278223585
2025-03-01T05:12:12Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
621
Amfibiyen
Amphibian
Biology
Amfibiyen Amfibiyenler, Amphibia sınıfını oluşturan, dıştan ısınan, amnionu olmayan, dört bacaklı omurgalı hayvanlardır. En geniş anlamda, amniyotları (modern sürüngenler, kuşlar ve memeliler gibi amniyon zarına sahip tetrapodlar) hariç tutan, tüm tetrapodları kapsayan bir parafiletik gruptur. Tüm mevcut (yaşayan) amf...
0
24
24
0
0
23,290
https://en.wikipedia.org/wiki/Amphibian
1,279,630,316
https://en.wikipedia.org/w/index.php?oldid=1279630316
2025-03-09T17:22:36Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
627
Tarım
Agriculture
Earth and environment
"Tarım\n\nTarım, gıda ve gıda dışı ürünler için bitkisel ve hayvansal üretim, su ürünle(...TRUNCATED)
0
20
20
0
0
18,825
https://en.wikipedia.org/wiki/Agriculture
1,280,911,579
https://en.wikipedia.org/w/index.php?oldid=1280911579
2025-03-17T04:00:54Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
633
Algler
Algae
Biology
"Algler\n\nAlgler (İngiltere: /ˈælɡiː/ AL-ghee, ABD: /ˈældʒiː/ AL-jee; tekil: alga /ˈælɡ(...TRUNCATED)
0
11
11
0
0
10,215
https://en.wikipedia.org/wiki/Algae
1,280,034,959
https://en.wikipedia.org/w/index.php?oldid=1280034959
2025-03-12T02:11:01Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
634
Varyans analizi
Analysis of variance
Mathematics
"Varyans analizi\n\nVaryans analizi (ANOVA), varyansı analiz ederek iki veya daha fazla grubun orta(...TRUNCATED)
114
11
11
0
0
9,872
https://en.wikipedia.org/wiki/Analysis_of_variance
1,274,337,899
https://en.wikipedia.org/w/index.php?oldid=1274337899
2025-02-06T19:43:54Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
639
Alkan
Alkane
Chemistry
"Alkan\n\nOrganik kimyada, bir alkan veya parafin (diğer anlamları da olan tarihsel bir basit ad),(...TRUNCATED)
0
11
11
0
0
11,313
https://en.wikipedia.org/wiki/Alkane
1,277,343,108
https://en.wikipedia.org/w/index.php?oldid=1277343108
2025-02-24T02:48:23Z
Qwen/Qwen3.8-27B@1d4bf0f (vllm 0.33.0, --quantization fp8, --max-model-len 8192; temperature 0)
extractor wh-1, chunker wp-2m, prompt wt-1, validator kd-4w4f
CC BY-SA 4.0
End of preview. Expand in Data Studio

Vikipedi-Bilim-TR: 300,218 science articles machine-translated into Turkish

Turkish machine translations of English Wikipedia science articles

300,218 English Wikipedia articles on biology, medicine, engineering, computing, space, the earth sciences, mathematics, chemistry and physics, machine-translated into Turkish with Qwen3.8-27B: 520 million Llama 3.2 tokens and 1.72 billion characters, one row per article. Formulas keep their original LaTeX between $ signs, and a translation that changed a formula was rejected. The English articles are licensed under CC BY-SA 4.0, and so is this dataset.

No person read the translations. Their quality was measured only with automatic checks against the English source and with a FLORES+ system score; both are described below.

Articles 300,218
Characters 1,723,795,123
Words (split on whitespace) 220,316,594
Llama 3.2 tokens 519,921,986
Translated chunks 643,977
Files 4 Parquet files, 0.73 GB
Topic Articles
Biology 118,005
Engineering 40,522
Medicine & Health 37,182
Computing 25,991
Space 22,251
Earth and environment 21,202
Mathematics 14,842
Chemistry 12,939
Physics 7,284

Columns and files

The articles are in data/train-*.parquet, sorted by page ID, in row groups of about 200 MB. reports/ holds the counts in this card as JSON: summary, article_selection, text_extraction, validation, article_gate, decontamination, personal_data and flores_plus.

Column Description
page_id English Wikipedia page ID
title Turkish title, the first line of text; null in the 64 articles whose first chunk failed the checks
title_en English title
topic the highest-scoring of the nine topics above
text Turkish translation of the article, starting with its title
formulas LaTeX formulas in the English text
num_chunks chunks the English article was cut into for translation
translated_chunks chunks whose translation passed the checks
edge_chunks_removed failed chunks at the start or end of the article that were cut
interior_gaps […] markers in text, each standing for one or more failed chunks
llama32_tokens tokens in text under the Llama 3.2 tokenizer
source_url the English article
source_revision, source_revision_url the revision that was translated and a permanent link to it
source_date_modified time of that revision
translator, pipeline, license model and server, pipeline versions and license; the same on every row
from datasets import load_dataset

ds = load_dataset("crudfab/vikipedi-bilim-tr", split="train", streaming=True)
article = next(iter(ds))
print(article["title"], article["text"][:500])

How the articles were chosen

Topics come from the English Wikipedia CirrusSearch dump of 29 December 2025 (enwiki-20251229-cirrussearch-content.json.gz, 7,113,605 articles). It stores the scores of Wikimedia's article topic classifier, from 1 to 1,000 per topic. An article was kept if the highest of its scores for nine topics (Physics, Chemistry, Mathematics, Biology, Space, Earth and environment, Medicine & Health, Computing, Engineering) was 700 or more; the other STEM labels (Technology, Libraries & Information and the general STEM label) were not used. The filters applied in this order:

Filter Articles left
At least one STEM topic tag 1,143,849
Not a list, index, outline or glossary page, by title 1,126,342
Not a disambiguation page, by category 1,075,747
Biography score under 500 979,211
No stub template 495,896
Highest of the nine topic scores 700 or more 372,591
Search text of 1,500 characters or more 312,089

Biographies were left out so that the dataset does not carry the personal details of living people. Most of the stubs were species articles.

The text comes from the Wikimedia Enterprise HTML dump of 20 March 2025 (enwiki-NS0-20250320-ENTERPRISE-HTML.json.tar.gz, 140,267,048,582 bytes, md5 2276dbb8db3bc93eabc90a505117e373), which holds the rendered article HTML. 304,132 of the selected articles are in it; the other 7,957 are not in that dump. The 3,053 articles in the category Living people were excluded, and 33 had no text left after extraction. 301,046 articles went to translation.

The HTML was converted to plain text: headings on their own line, paragraphs separated by a blank line, list items starting with "- ". Formulas keep their LaTeX source, inline as $…$ and displayed as $$…$$; chemical formulas become $\ce{…}$. Subscript and superscript digits become Unicode characters (H₂SO₄, x²). Footnote marks, the reference, notes, see-also and external-link sections, infoboxes, tables, figures and captions, navigation boxes, hatnotes and maintenance templates were removed. The editor names and IP addresses in the dump's revision fields were not read.

How they were translated

Each article was cut into chunks of up to 3,600 characters. The cut falls, in order of preference, at a line that ends in punctuation, at any line end or at a sentence end inside a line, never inside a word or a formula; a chunk could grow to 6,000 characters to reach such a point. On 2,000 extracted articles (4,188 chunks, 4,509 formulas), every boundary fell at a line or sentence end, none inside a word or a formula, and the chunks joined back into the original text. The 301,046 articles gave 647,881 chunks.

Each chunk was translated in one request, without the text of neighbouring chunks, by Qwen/Qwen3.8-27B at revision 1d4bf0f (Apache-2.0). Every server ran vLLM 0.33.0 with dynamic FP8 quantization and an 8,192-token context, and the driver checked each server's vLLM version, model and context length before sending it work.

  • System prompt: "You are a professional English-to-Turkish translator. Translate the passage into natural, fluent Turkish, faithfully and completely: do not add, omit, summarize or explain anything. Keep every line break, heading, list marker and table layout; keep code, formulas, URLs and numbers unchanged and translate the words in headings, list items and table cells. Output only the Turkish translation of the passage."
  • User message: "Translate the text in <passage> into Turkish.", a blank line, then the chunk between <passage> and </passage> tags.
  • Temperature 0, thinking disabled, at most min(4,096, max(768, 0.75 × source characters)) output tokens, a fixed seed per chunk.

A translation rejected by the checks below was translated once more with temperature 0.7, top_p 0.8 and top_k 20, adding presence_penalty 1.0 for repetition. After a rejection for untranslated text the request began with "Translate every line into Turkish, including headings, list items, table cells, captions and quotations. Do not leave any English in the output except names, code and URLs." After a formula rejection it began with "Copy every formula between $ signs exactly as it is, character for character and in the same order; translate only the words around the formulas." A translation that hit the token limit was retried once with the retry settings and presence_penalty 1.0; if it hit the limit again, the chunk was split in two at a boundary outside formulas, up to three times. Empty responses and server errors were retried up to five times.

The full run took place on 18 November 2026, after three pilot runs on the same day that set the checks.

Checks

Each translated chunk was compared with its English source. A chunk can fail more than one check. The language checks read only the translation of prose lines: source lines with at least two English function words that are not references, code, links, number tables or lists of names, because those correctly keep their English names and titles.

Check Rule Failed first attempt Still failed after retry
Formula changed a LaTeX formula of the source is missing or altered; whitespace, trailing punctuation, spacing commands, words inside \text{…} and braces around a single character are ignored 3,574 1,627
Untranslated under 4% Turkish-specific letters (ç, ğ, ı, ö, ş, ü) among 200 or more Latin letters and under 2% common Turkish function words, formulas, LaTeX commands and links not counted; or 100 or more letters of English prose copied unchanged 209 99
Too long over 1.3 times the length of a source of 500 or more characters 153 55
English left in 15% or more English function words among 50 or more words 94 30
Repetition repeated word 5-grams exceed the source's share by 0.15, or the text compresses 1/0.75 times better than the source 57 10
Cut short the source ends a sentence and the translation ends with a letter, comma or semicolon, unless the lines align and the last line is at least 80% as long as the source's 52 14
Too short under 0.8 times the length of a source of 500 or more characters 17 8
Prompt tag in the output <passage>, <context> and the like, when the source has none 1 0
Model preamble starts with "Türkçe çeviri:" or similar 0 0

Of 648,066 translated pieces (chunks and split halves), 3,985 (0.61%) failed at the first attempt and 1,805 (0.28%) still failed after the retry, nine in ten of them for a changed formula. No chunk was left without a translation: none exceeded the token limit after three splits and no request failed for good.

Failed chunks at the start or end of an article were cut. A failed chunk inside an article was replaced with a […] paragraph. An article was dropped if all its chunks failed, if the cut start and end exceeded 25% of its chunks, or if its failed interior chunks numbered more than max(1, 5% of all its chunks).

Article-level result
Articles dropped for failures at the start or end 461
Articles dropped because every chunk failed 142
Articles dropped for interior failures 124
Articles removed by the privacy filter that runs before publication 100
Article dropped because its title is an IP address ("1.1.1.1"), which would have been masked 1
Chunks cut at the start or end of kept articles 155
[…] markers in kept articles 607

The dropped articles are 0.275% of the translated articles and 0.541% of their characters; the limit, set before the run, was 2%.

Quality on FLORES+

On FLORES+ devtest English→Turkish (openlanguagedata/flores_plus at revision 1147e2d, 1,012 sentences), translated one sentence per request with the same prompts, model, server and temperature: chrF++ 57.43 and BLEU 26.99 (sacrebleu), COMET-22 90.87 (Unbabel/wmt22-comet-da). FLORES+ sentences come from Wikinews, Wikijunior and Wikivoyage; read the score as a system-level reference point. It does not measure how scientific terms were translated.

No chunk shares a 13-gram (Turkish, casefolded) with the FLORES+ Turkish dev and devtest sets, so the dataset can be used next to FLORES+ evaluations.

Personal data

Biographies were not selected, and articles in the category Living people were excluded. The articles still name scientists, physicians, engineers and other people in connection with their work. Strings shaped like an IP address (2,075 in 582 articles), a Turkish phone number (13) or a valid Turkish ID number (6), and e-mail addresses outside reserved example domains (191), were replaced with <IP>, <TELEFON>, <TCKN> and <EPOSTA>. In these articles the IP-shaped strings are software version numbers, dental formulas, vertex configurations of polyhedra, section numbers of standards and example addresses in networking articles, but the replacement was applied without exception. Enzyme Commission numbers written as "EC 3.1.3.5" have the same shape and were kept. To report content that should be removed, open a discussion on this dataset.

Limitations

  • No person reviewed the translations. Errors, omissions and literal renderings occur.
  • Wrong scientific terms are common, above all in species, anatomy, astronomy and physics vocabulary, and a negation is occasionally lost. Examples checked against the source: "phononic crystals" became "fotonik kristaller" (photonic), "the constellation Aquarius" became "Balıkçılık takımyıldızı" (Aquarius is Kova), "not dentate" became "dişlidir" (dentate) and "whiskers" became "sakallar" (beards). Established Turkish terms were not looked up.
  • Chunks were translated independently, so a term can be rendered differently in different parts of the same article.
  • Tables, infoboxes, figures and references were removed, so the text can mention a table, a figure or a citation that is not there.
  • […] markers interrupt the text where a chunk failed the checks.
  • The English text is the version of 20 March 2025.

License

This dataset is licensed under CC BY-SA 4.0, the license of the English Wikipedia text it was translated from. Attribution: the original articles were written by Wikipedia contributors; each row's source_revision_url links to the exact revision that was translated, and its page history lists the authors. Changes: the HTML was converted to plain text with references, tables, infoboxes, figures and navigation removed, the text was machine-translated into Turkish, and chunks that failed the checks were cut or replaced with […]. Redistributions and adaptations must keep this attribution and the CC BY-SA 4.0 license. Qwen3.8-27B is licensed under Apache-2.0, which places no restriction on its outputs. Wikipedia is a trademark of the Wikimedia Foundation; this dataset is not affiliated with or endorsed by it.

Türkçe özet

İngilizce Vikipedi'nin biyoloji, tıp, mühendislik, bilgisayar, uzay, yer bilimleri, matematik, kimya ve fizik maddelerinden 300.218 tanesinin Qwen3.8-27B ile Türkçeye makine çevirisi: 520 milyon Llama 3.2 token, madde başına bir satır, CC BY-SA 4.0 lisanslı. Formüller özgün LaTeX'iyle $ işaretleri arasında; formülü değiştiren çeviri reddedildi.

Maddeler Vikipedi'nin konu sınıflandırıcısının puanlarıyla seçildi: liste, anlam ayrımı, biyografi ve taslak sayfaları dışarıda; dokuz bilim konusundan birinde puan 700 ve üstü, metin 1.500 karakter ve üstü. Yaşayan kişiler kategorisindeki 3.053 madde çıkarıldı. Metin 20 Mart 2025 tarihli HTML dökümünden; kaynakça, tablo, bilgi kutusu ve şekiller atıldı.

Çeviriler insan tarafından okunmadı; kalite yalnız İngilizce kaynakla karşılaştıran otomatik denetimle ve FLORES+ sistem puanıyla ölçüldü (chrF++ 57,43, BLEU 26,99, COMET-22 90,87). Bilimsel terim hataları sık (tür, anatomi, gökbilim ve fizik terimleri); olumsuzluk ara sıra düşüyor. Maddeler en çok 3.600 karakterlik parçalarla, sıcaklık 0'da çevrildi; parça hiçbir zaman kelime ya da formül içinden kesilmedi. Denetimden geçemeyen parça bir kez yeniden çevrildi; yine kalan parça maddenin başında ya da sonundaysa kesildi, içindeyse yerine […] kondu. Kalan retlerin onda dokuzu değişen formül; düşen madde sayıca %0,275, karakterce %0,541. IP, telefon, kimlik numarası ve e-posta biçimindeki dizeler tipli yer tutucuyla değiştirildi; FLORES+ ile 13-gram örtüşmesi yok. Her satırda İngilizce maddenin adresi ve çevrilen revizyonun kalıcı bağlantısı var; atıf ve değişiklik bildirimi yukarıda. Bütün sayılar yukarıda ve reports/ altında.

Downloads last month
64

Collection including crudfab/vikipedi-bilim-tr