TR-OOLONG

A long-context aggregation benchmark for Turkish, with English counterparts built by the same pipeline. Each document joins thousands of real records (reviews, complaints, news articles, voice commands) into one text of 36K to 1M tokens, and each question asks about the whole collection:

Bu kayıtlarda hangisi daha çok: 'turizm' etiketli kayıtlar mı,
'magazin' etiketli kayıtlar mı?                              -> magazin  (44 vs 41 of 266)
Bu kayıtlarda kaç tane 'ulaşım' etiketli kayıt var?          -> 166

The label of a record is never written in the text, so a model has to decide what each record is about and then count. The core questions (the first example) compare two categories whose counts are so close that neither sampling part of the document nor searching for the relevant records answers them. Every answer is computed from the source dataset's own labels, twice, by independent code. The construction follows OOLONG (Bertsch et al., 2025).

Built with tr-oolong v0.10.0. Developed at the Institute for Data Science & Artificial Intelligence (DSAI), Boğaziçi University, as MSc thesis work. Full details: DATACARD.md in this repository.

At a glance

11 subsets · 347 documents · 2,362 questions · 102.5M tokens · 7 question types

subset	lang	classes	docs	questions	shortest	longest	max records in one doc
sikayet_tr	tr	29	49	366	99,861	1,016,268	10,000
interpress_tr	tr	16	43	337	99,091	998,305	2,941
amazon_hpc_en	en	3	43	237	98,426	987,208	17,087
sinema_tr	tr	3	38	212	99,176	720,375	6,000
marc_en	en	3	38	216	98,244	522,220	12,000
vitamins_tr	tr	3	38	187	99,223	496,474	12,946
musteri_tr	tr	3	38	213	88,854	496,144	13,825
tr_intent	tr	48	10	139	49,981	99,998	6,169
en_intent	en	48	10	133	49,921	99,871	8,122
tr_intent_paired	tr	48	20	161	47,629	99,057	6,000
en_intent_paired	en	48	20	161	36,250	75,187	6,000

Lengths are tokens under Qwen/Qwen3-8B. classes is the number of labels.

Question types and roles. Every question has a role:

role	question types	questions	what it shows
core	close_comparison: which are there more of, A or B? (both frequent, told apart from the text, counts very close)	309	that the model read and judged the whole document
retrieval	count with "rare": true (answer 5 to 30)	237	that it can find a few records by meaning
control	count 587, proportion 483, label_vs_label 299, most_common 156, least_common 147, second_most 144	1,816	that it can classify the records at all

Turkish/English pairs. Compare languages within a pair, never by totals (three subsets are Turkish only). tr_intent_paired and en_intent_paired contain the same utterances (MASSIVE is a human translation) in the same order, so all 161 questions, 41 of them core, are identical in both languages. tr_intent/en_intent match on token budget instead. musteri_tr/marc_en and vitamins_tr/amazon_hpc_en are different corpora with the same task; their core documents are built with identical sizes and identical positive/negative counts, so their 18 core questions each are identical too. In total 77 core questions are identical across the two languages. sikayet_tr, interpress_tr and sinema_tr are Turkish only.

Why the Turkish intent questions use English label names

In tr_intent and tr_intent_paired the question is Turkish but the label is MASSIVE's English identifier (transport_taxi, play_music). Translating the labels would put the answer back into the text: Turkish is verb-final, so a label like alarm_kur appears word for word in utterances such as "iki saat sonrasına alarm kur". On all 15,075 utterances, records containing their own label: English identifiers (shipped) 0.00%, Turkish imperative (müzik_çal) 3.13%, Turkish dictionary form (müzik_çalmak) 0.14%. The Turkish is in the text being classified; the label only names the bucket.

Relation to OOLONG

	OOLONG	TR-OOLONG
languages	English	Turkish, with English counterparts built the same way
document length	1K to 4M tokens (synthetic split)	36K to 1M tokens
labels per dataset	2 to 10	3, 10, 16, 29, 48
narrowing to a subset	to users or months printed on every record	none
questions over dates	yes, its hardest group	none yet (interpress_tr has dates)
same question, same answer in two languages	no	yes (*_intent_paired)
numeric score	partial (0.75 per unit of error)	partial, plus relative for large answers
questions that resist sampling and search	none found	309 close comparisons
published shortcut checks	none	yes, in the GitHub repository

OOLONG's construction code was not released; this is an independent implementation from the paper. OOLONG narrows questions to listed users or a month, both printed on every record, so a string search finds the relevant records: for user-narrowed questions it reads a median of 0.8% of the document and gets the exact answer 99% of the time. Its whole-document comparisons have a median gap of 38% between the two counts, so sampling answers them.

Loading and scoring

from datasets import load_dataset
qs = load_dataset("yigitates17/tr-oolong", "sikayet_tr", split="test")

• Join and pool on uid. id is unique only inside one subset.
• Give the model only the haystack text and the question.
• Score with src/scoring.py from the GitHub repository. It returns exact, partial (OOLONG's 0.75 ** |error|), relative (1 - |error| / answer) and primary, the one to report: exact for word answers (all core questions), partial for rare-label counts, relative for other numbers.
• Report the primary score per role, never pooled, and state how the model saw the document (one prompt, or an agent with code tools that can sample or search it).

Files

• questions.jsonl: uid, dataset, id, haystack_uid, haystack_id, language, target_tokens, kind, label / candidates / label_a / label_b, unit, answer, answer_key, rare, question, role, core_document.
• haystacks.jsonl (where the licence allows): uid, haystack_id, haystack, plus build metadata.
• manifest.json: seed, config, source hash, tokenizer, per-document lengths.

Subsets and licences

Each subset keeps the licence of its source.

subset	lang	source	license	text included	questions	note
tr_intent	tr	AmazonScience/massive (tr-TR)	cc-by-4.0	yes	139	MASSIVE is CC-BY-4.0: redistribution of derived data is permitted with attribution and a statement of changes.
en_intent	en	AmazonScience/massive (en-US)	cc-by-4.0	yes	133	As above; this is the parallel English twin.
tr_intent_paired	tr	AmazonScience/massive (tr-TR), record-matched	cc-by-4.0	yes	161	Record-matched with en_intent_paired: the same utterances in the same order, so all 161 questions (41 core) have the same answer in both languages (word answers via answer_key).
en_intent_paired	en	AmazonScience/massive (en-US), record-matched	cc-by-4.0	yes	161	The English half of the record-matched pair. Sized in records, not tokens, so its documents are shorter in tokens than the Turkish half (1.34x under the Qwen3-8B tokenizer).
vitamins_tr	tr	turkish-nlp-suite/vitamins-supplements-reviews (Vitaminler.com)	cc-by-sa-4.0	yes	187	CC-BY-SA-4.0 is SHARE-ALIKE: this subset and anything derived from it must stay CC-BY-SA-4.0. Cite Altinok (ACL 2023).
musteri_tr	tr	turkish-nlp-suite/MusteriYorumlari (Hepsiburada, Trendyol)	cc-by-sa-4.0	yes	213	CC-BY-SA-4.0 is SHARE-ALIKE: this subset and anything derived from it must stay CC-BY-SA-4.0. Labels are the customer's own 1-5 star rating.
marc_en	en	SetFit/amazon_reviews_multi_en (Multilingual Amazon Reviews Corpus)	apache-2.0	yes	216	Apache-2.0: redistribution permitted. The English half of the cleanest pair; labels are the reviewer's own 1-5 star rating.
amazon_hpc_en	en	McAuley-Lab/Amazon-Reviews-2023 (Health_and_Personal_Care)	other	no	237	Review text is governed by Amazon's Conditions of Use, not by the repository's license. Text withheld; rebuild locally with scripts/health.py + configs/amazon_hpc_en.json.
sikayet_tr	tr	Kaggle savasy/multiclass-classification-data-for-turkish-tc32	other	no	366	29 categories of Turkish consumer complaints. The uploader declares NO license and the text is scraped from a complaints site, so the text is withheld. Rebuild locally with scripts/sikayet_tr.py, which takes the path to a copy of ticaret-yorum.csv downloaded from Kaggle. Three of the original 32 categories were dropped because their names appear in over 30% of their complaints.
interpress_tr	tr	Interpress Turkish news category corpus, 270k	other	no	337	16 sections of Turkish news (17 in the source), with daily publication dates. The upstream card declares no license. Text withheld; rebuild locally with scripts/interpress_tr.py, which fetches the archive the Hugging Face loading script points at and verifies its sha256. The Apache header on that loading script covers the SCRIPT, not the data, and must not be cited as the data's license.
sinema_tr	tr	turkish-nlp-suite/BuyukSinema	cc-by-sa-4.0	yes	212	Turkish film reviews; sentiment from the reviewer's own 10-point rating (1-4 negative, 5-6 neutral, 7-10 positive; exact ratings cannot be read from the text). Share-alike: anything derived from this subset stays CC-BY-SA-4.0.

Subsets without text

• amazon_hpc_en (McAuley-Lab/Amazon-Reviews-2023 (Health_and_Personal_Care)) -- Review text is governed by Amazon's Conditions of Use, not by the repository's license. Text withheld; rebuild locally with scripts/health.py + configs/amazon_hpc_en.json.
• sikayet_tr (Kaggle savasy/multiclass-classification-data-for-turkish-tc32) -- 29 categories of Turkish consumer complaints. The uploader declares NO license and the text is scraped from a complaints site, so the text is withheld. Rebuild locally with scripts/sikayet_tr.py, which takes the path to a copy of ticaret-yorum.csv downloaded from Kaggle. Three of the original 32 categories were dropped because their names appear in over 30% of their complaints.
• interpress_tr (Interpress Turkish news category corpus, 270k) -- 16 sections of Turkish news (17 in the source), with daily publication dates. The upstream card declares no license. Text withheld; rebuild locally with scripts/interpress_tr.py, which fetches the archive the Hugging Face loading script points at and verifies its sha256. The Apache header on that loading script covers the SCRIPT, not the data, and must not be cited as the data's license.

For these, the text is rebuilt locally. The build is deterministic, so the result is byte-identical:

git clone https://github.com/yigitates17/tr-oolong && cd tr-oolong
python scripts/<fetch_script>.py
python src/build_tr_oolong.py --config configs/<set>.json --build

What each role measures, and limitations

No model has been run on this benchmark yet. The statements below come from simulated readers: short programs that are told the true label of every record they read, so they show what a reading strategy can achieve.

role	strongest shortcut found	reader of everything
core	sampling half the document 0.62; the 10% most relevant records by topic search 0.54; guessing 0.50	0.86 at 95% labelling accuracy, 0.76 at 90%
retrieval	topic search reading 5%: 0.53 under partial	depends strongly on labelling accuracy
control	sampling 5%: about 0.8 under relative	about 0.9 or more

• Every release also passes four checks: searching the text for label names (0 of 856,798 shipped records contain one), always giving the most common answer, answering from the source corpus's label shares, and guessing labels from record length and punctuation.
• Control questions are not evidence of reading: a 5% sample answers them almost as well as the whole document, and under relative a longer document is not harder.
• 309 questions are core (margin of error about ±0.06 on a model's core score). They only compare label pairs that can be told apart from the text (lists in the GitHub repository). The 214 from core documents (core_document: true), built from exact label counts with which label is larger and which is named first balanced by design, are the cleaner group: first-named 0.50, source-dataset shares 0.50, topic search 0.49. The 95 from ordinary documents are slightly exposed (source-dataset shares 0.62, topic search 0.67). Report the two groups separately.
• Cleaning (v0.10.0): newspaper mastheads, reviews that write their score ("7/10", "5 stars") and HTML are removed.
• Wrong labels in the source data decide some close comparisons, so even a perfect model scores below 1.0 on core questions (about 0.85 with 5% of labels wrong).
• Brand questions (in v0.7.x) were removed in v0.8.0: searching for the printed brand answered all of them. The v0.7.x per-question difficulty grades were withdrawn as well.
• Questions on one document overlap (337 are a count and a proportion of the same thing); treat the document as the unit of evidence.
• Documents of the same subset and length share 20-39% of their records.
• Intent labels: 2.7-9.3% judged wrong on a 150-record check, partly from translation, which affects only the Turkish half.
• Every record comes from a public dataset; a model that memorised a dataset's labels could label records without reading them (not tested).
• No time-based questions. Lengths use one tokenizer; the Turkish/English token ratio on the same sentences ranges from 0.57x to 2.16x across tokenizers.

Citation

@misc{troolong,
  title  = {TR-OOLONG: A Turkish Long-Context Aggregation Benchmark},
  author = {Ate{\c{s}}, Yi{\u{g}}it},
  year   = {2026},
  note   = {Bo{\u{g}}azi{\c{c}}i University, Institute for Data Science \& Artificial Intelligence},
  url    = {https://github.com/yigitates17/tr-oolong}
}

Please also cite OOLONG (Bertsch et al., 2025, arXiv:2511.02817) and the source corpus of each subset you use.