imrunya/en-tr-parallel-filtered-25m

English–Turkish filtered parallel corpus (25.5M pairs)

25,464,091 English–Turkish sentence pairs from eight public corpora, filtered, de-duplicated and cleared of every sentence of the usual test sets. Built for training English→Turkish translation models from random initialisation (Qwen3-0.6B architecture, autoregressive and diffusion).

Columns

column	meaning
en, tr	English and Turkish sentence
corpus	source corpus (see below)
labse	LaBSE cosine similarity of the pair (every pair is ≥ 0.72)
tier	smallest nested training tier that contains the pair: s (253,561 pairs), m (1M), l (4M), xl (all)

The tiers are nested and the rows are in tier order: tier m is every row with tier s or m (the first 1,000,000 rows), and so on. Each tier is filled with the most reliable data first: human translations, then web-crawled, then mined pairs.

from datasets import load_dataset
ds = load_dataset("imrunya/en-tr-parallel-filtered-25m", split="train")
tier_m = ds.select(range(1_000_000))          # or ds.filter(lambda r: r["tier"] in ("s", "m"))

The validation split holds 2,000 held-out pairs from the same pool (not used for training).

Sources

corpus	pairs kept	of scanned	kind
ccmatrix (CCMatrix)	22,484,463	47,045,956	mined from web text with LASER
macocu (MaCoCu-tr-en 2.0)	985,444	1,626,006	web-crawled, document-aligned
akademik (YÖK thesis abstracts)	942,340	1,399,159	abstracts written in both languages
wikimatrix (WikiMatrix v1)	324,449	1,486,306	mined from Wikipedia
ted (TED2020)	292,789	367,612	human-translated talk subtitles
covost (CoVoST 2 en→tr)	244,045	320,492	human translations of Common Voice sentences
setimes (SETIMES2)	186,339	207,678	human-translated news
globalvoices (Global Voices)	4,222	5,481	human-translated news

How it was filtered

In this order, for every pair:

1. Rules: length and character-ratio limits, digit mismatch between the sides, URLs, copies of the source, mojibake, Turkish written without Turkish letters, exact duplicates.
2. Test-set exclusion: any pair whose source or target matches a test sentence exactly, after case and punctuation normalisation, or shares a 13-word sequence with one. Test sets: FLORES-200 dev and devtest, WMT24++ en-tr, WMT newstest 2016–2018, NTREX-128.
3. Language identification (lingua, high-accuracy mode, 25 candidate languages) on both sides.
4. Meaning similarity: LaBSE cosine ≥ 0.72.
5. Sentence-shift test: drops pairs where one side carries an extra sentence.
6. Cross-corpus de-duplication on pair, source and target.
7. Length: at most 256 source and 192 target tokens with the Qwen3 tokenizer.

Known noise (from a manual read of samples and automatic checks)

• 2–8 % of pairs per corpus have LaBSE below 0.75, i.e. close to the threshold; these are the loosest translations (most in akademik, ted, ccmatrix).
• ccmatrix pairs are mined, not translated: some differ in a name or number (seen in a sample: "Vietnam" paired with "Mısır", Egypt) or drop part of the sentence.
• setimes: 2.8 % are bylines and source lines ("By … for Southeast European Times …"); some translations add a subject that is only implied in English.
• ted: 4.3 % contain stage notes such as "(Laughter)"; subtitle segments sometimes lack final punctuation.
• akademik: thesis headers with names and committee lists appear as pairs and can be misaligned.
• Sentence-final punctuation differs between the sides in 2–16 % of pairs (highest in covost, where English sentences are often quoted).

Licence

This is a compilation; each pair keeps the terms of its source corpus (see the corpus column). Several sources restrict reuse (for example TED talks are CC BY-NC-ND, WikiMatrix is CC BY-SA, and CCMatrix text comes from web pages whose rights stay with their authors). Check the terms of each source before redistributing or using the data commercially.

Results with this data

Qwen3-0.6B architecture trained from random weights, FLORES-200 devtest en→tr, chrF++ / COMET-22 (AR, beam 5): 43.6 / 75.3 (tier s), 53.3 / 85.6 (m), 57.5 / 87.9 (l), 59.0 / 88.4 (xl).