TR-OOLONG

A Turkish long-context aggregation benchmark with a matched English twin, built by an identical pipeline. Questions ask distributional facts about a 50K-1M-token haystack ("how many records are labeled X?", "which label is most common?"); every gold answer is computed exactly from the source labels by two independent code paths, so there is no manual annotation and nothing is grep-solvable.

Developed at the Institute for Data Science & Artificial Intelligence (DSAI), Boğaziçi University, as MSc thesis work.

Built with tr-oolong v0.7.0. See that repository for the builder, the configs that reproduce every set byte-for-byte, and DESIGN_DECISIONS.md for why each construction choice was made.

At a glance

11 subsets · 195 documents · 2,240 questions · 50.7M tokens · 2 languages · 9 question families

Every question carries a measured difficulty grade in difficulty.jsonl: how well the best of four partial readers does on that question alone. 259 questions (11.6%) are graded very hard, 1,609 (71.8%) easy. Read the section on it below before reporting any score.

subset	lang	classes	docs	questions	shortest	longest	max records in one doc
sikayet_tr	tr	29	25	300	99,861	999,625	9,943
interpress_tr	tr	16	25	300	99,267	998,392	2,890
amazon_hpc_en	en	3	25	292	98,559	987,623	16,116
sinema_tr	tr	10	20	240	99,179	497,150	4,057
vitamins_tr	tr	3	20	236	99,204	496,526	12,898
musteri_tr	tr	3	20	199	99,137	496,238	13,618
marc_en	en	3	20	193	98,232	491,821	11,080
tr_intent	tr	48	10	120	49,981	99,998	6,169
en_intent	en	48	10	120	49,921	99,871	8,122
tr_intent_paired	tr	48	10	120	47,629	99,057	6,000
en_intent_paired	en	48	10	120	36,250	75,187	6,000

Lengths are tokens under Qwen/Qwen3-8B. Every document also records n_chars, so lengths can be re-derived under a different tokenizer without rebuilding — which matters, because the Turkish/English token ratio on identical content runs from 0.57x to 2.16x depending on whose tokenizer counts it.

Question families (2,240 total): count 921 · proportion 492 · label_vs_label 292 · most_common 138 · least_common 132 · second_most 128 · entity_count 87 · pairwise 35 · entity_argmax 15.

265 of the 921 counts are rare-label counts (v0.7.0): the answer holds 5 to 30 records. They are worded identically to any other count and carry "rare": true. They exist because a small answer is the only thing that resists a partial reader: a reader that opens nothing and answers N/K scores 0.43-0.55 on an ordinary count and 0.000 on a rare one, and a corpus-prior oracle falls from 0.41-0.50 to 0.06-0.12.

What the questions look like

Real questions from the release, with their gold answers:

[count/tr]         Bu yorumlardan kaç tanesi 'olumlu' etiketli?              -> 1600
[proportion/tr]    Yorumların yüzde kaçı 'olumlu' etiketli?                  -> 62
[most_common/tr]   Bu yorumlarda en sık görülen etiket hangisi?
                   Etiketler: 'nötr', 'olumlu', 'olumsuz'.                   -> olumlu
[entity_count/tr]  'Venatura' markası hakkındaki yorumlardan kaç tanesi
                   'nötr'?                                                    -> 10
[pairwise/tr]      'olumlu' yorumu hangisinde daha çok: 'Shorne' mi yoksa
                   'Tab' mı?                                                  -> Tab
[shift/tr]         Yorumların ikinci yarısında 'olumsuz' oranı ilk yarıya
                   göre arttı mı azaldı mı?                                   -> azaldı

[least_common/en]  Which label is the least common in these records?
                   Labels: 'alarm_set', 'lists_createoradd', 'music_query',
                   'play_audiobook', 'qa_currency'.                          -> play_audiobook
[label_vs_label/en] Are records labeled 'datetime_query' more common, less
                   common, or the same frequency as 'audio_volume_up'?       -> the same

None of these answers appears anywhere in the text. The label is latent: a model has to decide what each record means before it can count anything. Records whose text contains any label's surface form are dropped at build time, so a substring search returns nothing useful.

Why the Turkish intent questions name English labels

This is deliberate, not an oversight. In tr_intent and tr_intent_paired the question is Turkish but the label is the source corpus's English identifier (transport_taxi, play_music), because translating the labels into Turkish puts the answer back into the text. Turkish is verb-final, so a noun_verb label reproduces a natural Turkish phrase: the label alarm_kur appears verbatim inside utterances like "iki saat sonrasına alarm kur".

Measured over the full 48-label space on the same 15,075 utterances:

labels used	records leaking their own label
English identifiers (what ships)	0.00%
Turkish, imperative form (müzik_çal)	3.13% (472 records)
Turkish, dictionary form (müzik_çalmak)	0.14% (21 records)

Keeping the English identifiers loses no Turkish signal, because the Turkish is in the text being classified — the label is only the name of the bucket. The translated variants exist in the repository under configs/experimental/ for anyone who wants to study the trade-off, and are deliberately not part of this release.

The matched twin

tr_intent_paired and en_intent_paired contain the same utterances, in the same order, with the same labels — one is the translation of the other. So the same question has the same correct answer in both languages:

TR: Bu kayıtlarda kaç tane 'transport_taxi' etiketli kayıt var?   -> 18
EN: How many utterances have the intent 'transport_taxi'?          -> 18

100 of 120 question pairs share a byte-identical gold answer; the other 20 (10 shift, 10 label_vs_label) state the same fact in language-specific strings (arttı / rose, eşit / the same), so all 120 are paired. A score difference between the two halves is therefore not a property of the question. It can still come from the language, from the translation (the Turkish half is a human localization, and part of its measured label noise is mistranslation), or from the Turkish text costing 1.3x the tokens under the reference tokenizer. The two halves can be compared with a paired test.

tr_intent / en_intent are the same corpus matched on token budget instead of record count — so the two halves hold different numbers of records and their answers do not correspond. That pair asks "at equal cost"; the paired sets ask "at equal content".

Relation to Oolong

This follows the construction principle of Oolong (Bertsch et al., 2025) and extends it. Their construction code was unreleased at the time of writing, so the pipeline here is an independent reimplementation from the paper's description.

	Oolong	TR-OOLONG
languages	English	Turkish + a matched English twin
documents	not reported per split	195, 50.7M tokens
context length	reported at 8K-128K	36K-1.0M
label space	2-10 classes	3, 10, 16, 29 and 48
grouping axis	synthetic user IDs	real brands
timeline questions over real dates	6 families	✗ none built — interpress_tr ships real dates, but no family uses them yet
cross-lingual	✗	✓ same question, same answer, two languages
numeric metric	0.75^|y-yhat|	same + a scale-free one
shortcut audit	not reported	5 solvers, reports shipped

Where Oolong is harder: it has six question families conditioned on real calendar dates, which its paper reports as its hardest group. There is no equivalent here, because no Turkish labelled corpus with dates was found. The substitute — comparing the first half of a document to the second — is binary and is the weakest family in this release.

Where this is harder: 48 classes against their 2-10, documents to 988K tokens, and answers in the thousands where theirs are single digits. That last difference is not purely an advantage — see Limitations.

Subsets and their licenses

Each subset carries the license of its source corpus. They differ. Read the row for the subset you use.

subset	lang	source	license	text included	questions	note
tr_intent	tr	AmazonScience/massive (tr-TR)	cc-by-4.0	yes	120	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	120	As above; this is the parallel English twin.
tr_intent_paired	tr	AmazonScience/massive (tr-TR), record-matched	cc-by-4.0	yes	120	RECORD-MATCHED twin: the same utterances, in the same order, as en_intent_paired. 100 of 120 questions share a byte-identical gold answer and the other 20 the same fact in language-specific strings, so the two can be compared with a paired test.
en_intent_paired	en	AmazonScience/massive (en-US), record-matched	cc-by-4.0	yes	120	The English half of the record-matched pair. Sized in RECORDS, not tokens, so its token counts are lower than the Turkish half by the morphology factor (1.30-1.34x).
vitamins_tr	tr	turkish-nlp-suite/vitamins-supplements-reviews (Vitaminler.com)	cc-by-sa-4.0	yes	236	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	199	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. No entity column, so six families ship.
marc_en	en	SetFit/amazon_reviews_multi_en (Multilingual Amazon Reviews Corpus)	apache-2.0	yes	193	Apache-2.0: redistribution permitted. The English half of the cleanest pair; labels are the reviewer's own 1-5 star rating, and the entity families are omitted to stay parallel with musteri_tr.
amazon_hpc_en	en	McAuley-Lab/Amazon-Reviews-2023 (Health_and_Personal_Care)	other	no	292	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	300	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 for category-name leakage above 30%. Resolve the license before relying on this set.
interpress_tr	tr	Interpress Turkish news category corpus, 270k	other	no	300	17 categories of Turkish news 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	240	Turkish film reviews labelled with the reviewer's own 10-point rating. The only large-label-space Turkish source found with a declared license, so unlike the other two v0.7.0 additions its text ships normally. Share-alike: anything derived from this subset stays CC-BY-SA-4.0.

Subsets shipped without haystack 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 for category-name leakage above 30%. Resolve the license before relying on this set.
• interpress_tr (Interpress Turkish news category corpus, 270k) -- 17 categories of Turkish news 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, questions.jsonl and the manifest are included but the haystack text is not, because the source license does not permit redistributing it. Rebuild locally -- the build is deterministic, so you get byte-identical haystacks:

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

Fields

questions.jsonl -- one question per line:

field	meaning
id	unique question id
haystack_id	which haystack it refers to
language	tr or en
target_tokens	length tier of the haystack
kind	question family
label / entity / candidates	what the question is about
answer	gold answer, computed from source labels
answer_key	language-neutral form of the answer, where the answer is a word (label_vs_label). answer is what to score; this is for comparing the matched pair across languages
rare	present and true on a rare-label count: the gold answer is 5-30 records
question	the prompt text, self-contained

difficulty.jsonl -- one row per question, joined on id:

field	meaning
shortcut_score	best score any of the four partial readers achieved on this question at a 5% budget
shortcut_reader	which reader achieved it
difficulty	very hard / hard / moderate / easy, from that score
blind_score	what a reader that opens nothing scores
grade_se	standard error of the grade over 200 samples
borderline	true when the grade is within two standard errors of a band boundary and could flip

haystacks.jsonl -- one haystack per line: haystack_id, n_examples, drift_target, and haystack (the concatenated text). Only haystack and the question go to the model. n_examples and drift_target are build metadata for auditing; drift_target names the label the shift question asks about, so passing it into a prompt would hand the model half of that question. scripts/run_eval.py sends the text alone.

Scoring

Use src/scoring.py from the repository. It reports exact, partial (0.75**|y-yhat|, matching Oolong) and relative (scale-free). Do not re-implement it; the metric is frozen.

Report relative as lift over the blind reference, and state the reading protocol. Under relative a reader that opens nothing, counts the records and answers N/K already scores 0.43-0.63 on count and proportion. The per-family reference is in manifests/sampling_audit.json in the repository. State whether the model was given the document in a single prompt (scripts/run_eval.py) or run agentically with tools or code execution over it, and whether it was permitted to sample; these are different benchmark conditions and score very differently (see Limitations).

Limitations

Five shortcut solvers are run against every build. Four fail, as intended: substring search over label names, always answering the most frequent label, answering from corpus statistics without opening the haystack, and classifying records from length and punctuation alone. Their reports ship in the repository under manifests/.

The fifth partly succeeds, and it bounds what this benchmark shows. Because gold answers are large (median count near 1,000), most question families can be answered by classifying a part of the records and scaling up rather than by reading all of them. Measured with solvers given the true label of every record they read -- upper bounds, not model results -- a 5% random sample scores 0.89-0.92 on count and 0.98 on most_common on the review sets, against a read-nothing reference of 0.43-0.63 and 0.55. A reader that classifies only 5% of the records, half at the start of the document and half at the end, scores 0.90-0.91 on count and 1.00 on most_common: the same as random sampling, for a budget a truncating model already has. Reading the first 5% contiguously scores 0.65-0.70, but that is the one reading pattern the document layout penalises and it is not a defence. The 48-class intent axis resists better (0.45 on most_common) because its decision margins are narrower.

Under relative, the length axis is flat. A fixed budget of 1,000 randomly read records scores 0.94-0.97 on count at every tier from 100K to 1M tokens, because a proportion's error depends on how many records were read, not on how many exist. On the numeric families (64% of the questions) a relative score therefore cannot tell a model that read 1,000 records from one that read 16,000. What the length axis still tests is whether a model survives ingestion at all. A single relative score also cannot say whether a model read more or classified better: a perfect classifier reading 5% outscores a 90%-accurate classifier reading everything.

The shift family is withdrawn as of builder v0.7.0 and should not be scored. It asked whether a label's share rose or fell between the two halves of the document. A reader classifying fifty records at each end answers it perfectly: 1.000 on all eight subsets at a 25% budget, 0.90-1.00 at 5%, against a majority baseline of 0.50-0.70. The answer is a step function at a known position and its direction is one bit. shift questions are present in this published data. Discard them rather than caveating them; they will not be rebuilt.

Every question carries a measured difficulty grade, in difficulty.jsonl. Each question was run against four partial readers at a 5% budget; the grade is the best score any of them achieved on that question alone.

grade	questions	share	what it measures
very hard (< 0.35)	259	11.6%	whether the model aggregated over the document
hard	140	6.2%
moderate	232	10.4%
easy (>= 0.80)	1,609	71.8%	whether the model can classify Turkish records

Report the bands separately and report the gap between them; do not pool all 2,240 questions into one number. The gap estimates how much of the document a model read. easy 0.90 / very-hard 0.30 is a sampler. 0.40 on both means the model cannot classify Turkish, and its long-context result says nothing about long context.

Caveats that travel with the grade: it is relative to these four readers, which are given the true label of every record they read and are therefore upper bounds; it is a property of the question AND the proportional metric; and musteri_tr has 0 very-hard questions while marc_en has 1, so neither is evidence of aggregation difficulty.

So the supported claim is that this benchmark requires classifying latent Turkish labels and aggregating them. It does not establish that a model has processed the entire document, and under relative it does not establish that a longer document was harder. Exact-match scoring is immune to this; relative is not.

Other limitations, with numbers, are in DATACARD.md: the twelve questions on each 3-class document carry about two degrees of freedom (the unit of evidence is the document, not the question), the 750K tier of vitamins_tr is prior-exposed on the numeric families (the corpus-share guess scores 0.75 there), label noise as a per-family ceiling (measured 2.7-9.3% on the intent axis, part of it mistranslation and asymmetric across the twin), documents within a length tier sharing 21-39% of their records, no timeline axis, entity families being available only at 500K tokens and above, small per-family sample sizes, and all lengths measured under a single tokenizer.

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 the source corpus of whichever subset you use, and Bertsch et al. (2025), Oolong, arXiv:2511.02817, whose construction principle this follows.