atahanuz/setimes-en-tr-aligned-corpus

SETimes EN-TR — Sentence-Aligned, LLM-Cleaned

A cleaned and re-aligned version of the SETimes English-Turkish parallel corpus. The original SETimes data is paragraph-style — each "pair" can contain a headline, a dateline, several body sentences, and a source citation, all glued together on one line. This version splits everything into proper sentence pairs so each row is one English sentence next to its Turkish translation.

144,064 sentence pairs, split into train (142,064), held-out dev (1,000), and held-out test (1,000). The dev and test splits are the validation and evaluation sets used to benchmark translation models on this corpus; both are excluded from training to avoid leakage. (One pair that was duplicated across two source paragraphs was de-duplicated when carving out the dev split, hence 144,064 vs. the original 144,065.)

Quick stats

Stat	Value
Total pairs	144,064 (train 142,064 / dev 1,000 / test 1,000)
EN words per pair (median / mean)	20 / 20.7
TR words per pair (median / mean)	16 / 17.1
Similarity score (mean / p05 / p95)	0.80 / 0.70 / 0.87
Source paragraphs (before alignment)	119,162

Columns

Column	Type	Description
en	string	English sentence
tr	string	Turkish sentence
sim	float	Cross-lingual cosine similarity between the two sides, from qwen3-embedding-8b. Higher is more lexically similar — but note this reflects cognate overlap, not translation quality. Citation lines and dates score highest; valid translations of low-cognate sentences can score around 0.5.
source_line	int	Index of the original SETimes paragraph this sentence came from. Multiple rows can share the same source_line if the paragraph was split into several sentence pairs.

How it was cleaned

The original SETimes corpus has a few practical problems for training translation models:

1. Many "pairs" are not really one sentence — they're a whole news article header glued together: headline, date, location dateline, one or two body sentences, and a source citation, all on one line. Models that train on this learn long-distance patterns that don't generalize.
2. Some "pairs" are pure citations like (Reuters, AP - 12/04/05) — same on both sides, no translation signal.
3. A small number of pairs have wrong content on one side because of editorial revisions to the source articles.

To fix this, every paragraph pair was passed through a large language model (gpt-oss-120b) with a strict prompt that asked it to:

• Keep every full sentence and every headline (keeping the date next to the headline).
• Drop standalone datelines like ZAGREB, Croatia --, source citation blocks like (Reuters, AP - ...), and lone date stamps.
• Split body sentences off the headline when they were fused together by a comma — this case was a real problem in earlier versions of the prompt.
• Never translate, paraphrase, or correct typos — the model only chooses what to keep and how to split.

After the LLM step, every output sentence was embedded with qwen3-embedding-8b and pairs were filtered out if:

• The English and Turkish sides were identical (untranslated).
• The word-length ratio was extreme (outside [0.25, 4.0]).
• Either side was shorter than 3 words.
• The word overlap between the two sides was above 80% (cognate-heavy or copied).
• The cross-lingual cosine similarity was below 0.35 (the LLM had paired the wrong sentences across the two sides).

That removed 179 pairs out of 144,244 — a small but worthwhile cleanup.

How to load

from datasets import load_dataset
ds = load_dataset("atahanuz/setimes-en-tr-aligned-corpus")
print(ds)            # DatasetDict with 'train' (142,064), 'dev' (1,000), 'test' (1,000)
print(ds["train"][0])
# {'en': '...', 'tr': '...', 'sim': 0.73, 'source_line': 59}

Or directly from pandas:

import pandas as pd
train = pd.read_csv("setimes_en_tr_aligned_train.csv")
dev   = pd.read_csv("setimes_en_tr_aligned_dev.csv")
test  = pd.read_csv("setimes_en_tr_aligned_test.csv")

Source and license

The underlying text comes from the SETimes parallel corpus (Tyers and Alperen, 2010), a collection of news articles on Southeast European affairs published 2003-2012 by SETimes.com. SETimes is distributed under CC-BY-4.0 and this derivative keeps the same license.

If you use this dataset, please cite the original SETimes paper:

Tyers, F. M., & Alperen, M. S. (2010). South-East European Times: A parallel corpus of Balkan languages. Proceedings of the LREC workshop on exploitation of multilingual resources and tools for Central and (South-) Eastern European Languages.