--- license: - cc-by-sa-4.0 - cc-by-4.0 - other language: - tr - en pretty_name: TR-OOLONG size_categories: - 1K **Status: provisional, data only.** This release contains the documents, > questions, answers, label maps, construction rules, hashes and a standalone > script that recomputes every answer. The generation code will be published on > GitHub together with the paper; until then every answer can be checked, but > the documents cannot be regenerated by others. Until the paper, file formats > may still change; the version number will change with them, and files are > never changed in place. TR-OOLONG tests whether a model can aggregate over a long document: thousands of short labelled records (reviews, voice-assistant requests) are joined into one document, and each question asks about label counts over the whole document, over some users, or over some dates. It follows the design of OOLONG-synth (Bertsch et al., 2025) and adds Turkish: Turkish review sets, a Turkish-English parallel set, and a close-comparison extension where two label counts differ by 0-10%. ## Sets | set | lang | labels | licence | text | main docs / questions | lengths (tokens) | label agreement* | |---|---|---|---|:---:|---|---|---| | `musteri_tr` | tr | olumlu, olumsuz | CC-BY-SA-4.0 | yes | 20 / 500 | 2K-1M | 100% | | `sinema_tr` | tr | olumlu, olumsuz | CC-BY-SA-4.0 | yes | 24 / 600 | 2K-4M | 88% | | `massive_tr` | tr | 11 scenario groups | CC-BY-4.0 | yes | 16 / 400 | 1K-128K | 79% / 84%** | | `massive_en` | en | 11 scenario groups | CC-BY-4.0 | yes | 16 / 400 | 1K-128K | 83% | | `marc_en` | en | positive, negative | MARC terms | **no** | 24 / 600 | 2K-4M | 100% | | `amazon_hpc_en` | en | positive, negative | none declared | **no** | 22 / 550 | 2K-2M | 100% | | `vitamins_tr` | tr | olumlu, olumsuz | CC-BY-SA-4.0 | yes | 20 / 500 | 2K-1M | 92%, outside main results | *A native speaker labelled records blind and was then shown the source label; agreement is on records where the reviewer could decide. Turkish 140 records, English 80, 20-40 per set, then 72 fresh Turkish records to choose the MASSIVE and complaint groupings (the complaint set was then dropped). **MASSIVE groups on fresh records: 26/33 = 79% under the rule fixed in advance; the `general` scenario (jokes) was removed after that check, and on the 31 records left the groups reach 26/31 = 84%. `vitamins_tr` is kept outside the main results because 40% of its reviews state no opinion about the product. `massive_tr` and `massive_en` are the same documents in two languages (same records, order, users, dates and answers). Main set: 142 documents and 3,550 questions in total (MASSIVE's documents counted once per language), 25 per document, in three groups (counting, user, timeline). Lengths are powers of two from 1,024 tokens; a set stops where its records run out. **Close-comparison extension:** one question per document ("is label A more common, less common, or the same frequency as label B?") at margins of 0, 0.1, 0.5, 2 and 10% of the records, with 2,048, 8,192 or 32,768 records. 20 test documents per cell (50 cells), 4 development documents for the cells the development split can fill (15 cells). Every document has a labels-hidden and a labels-shown rendering of the same text. ## Files ``` data/main//test/{documents,questions}.jsonl.gz, manifest.json data/extension//{test,dev}/{documents,questions}.jsonl.gz, manifest.json CONSTRUCTION.md how documents, questions and answers are built label_maps.json source label -> released label provenance.json sources, revisions, file hashes, tokenizer, seeds recompute_answers.py recomputes every answer; restores withheld text restore_hints.json which source line to use for 5 texts repeated with other ratings SHA256SUMS hash of every file in this release ``` **Questions:** `uid`, `doc_id`, `language`, `question`, `answer` (a list; any listed answer is correct), `question_en`, `answer_en` (English original), main set also `task_group`, `task`, `answer_type`, `input_subset`; extension also `n_records`, `margin_target`, `margin_realized`, `tie`. **Documents:** `doc_id`, `source`, `language`, `split`, `n_records`, `n_tokens` (Qwen3-8B), `label_counts`, `seed`, `text` (labels hidden), `text_with_labels`; main set also `context_len` (target length), `temperature`, `attempt`; extension also `pair`, `winner`, `named_first`, `count_difference`, `margin_target`, `margin_realized`. Give the model `text` followed by `question`. `text_with_labels` is for analysis (or for a labels-shown condition), not for the labels-hidden task. Date answers are stored as `YYYY-MM-DD`; the questions ask for MM/DD/YYYY (Turkish: GG.AA.YYYY), so parse dates before comparing. ```python from datasets import load_dataset q = load_dataset("yigitates17/tr-oolong", "musteri_tr", split="test") d = load_dataset("yigitates17/tr-oolong", "musteri_tr_documents", split="test") ``` ## Sets without text The review text of `marc_en` and `amazon_hpc_en` may not be redistributed, so their documents ship without it: each keeps its header, closing lines, every record's date and user, a 16-hex-digit SHA-256 prefix of every record's text, and the SHA-256 of both renderings. Questions, answers and label counts are complete. To restore the text, download the two source files listed in `provenance.json` and run: ``` python recompute_answers.py data --restore marc_en=train.jsonl \ --restore amazon_hpc_en=Health_and_Personal_Care.jsonl --write restored ``` ## Licences and attribution Each set keeps its source's licence. - `musteri_tr`, `sinema_tr`, `vitamins_tr`: CC-BY-SA-4.0 (share-alike: these files and anything derived from them stay CC-BY-SA-4.0). Source: Turkish NLP Suite (Altinok, ACL 2023). Changes: cleaned, ratings mapped to two labels, records arranged into documents with synthetic users and dates. - `massive_tr`, `massive_en`: CC-BY-4.0. Source: MASSIVE (FitzGerald et al., 2023). Changes: scenarios grouped into 11 labels, `general` removed, utterances arranged into documents with synthetic users and dates. - `marc_en`: the Multilingual Amazon Reviews Corpus is licensed by Amazon for academic research only and may not be republished; it has been withdrawn from distribution by its provider. This release contains no MARC text, only questions, answers, counts and hashes. **Correction:** earlier versions of this dataset, up to v0.10.0, included MARC text, labelled Apache-2.0 after a third-party copy; that licence attribution was wrong. The text was withdrawn on 12 October 2026. Copies downloaded before then may still exist; please delete them. - `amazon_hpc_en`: Amazon Reviews 2023 declares no licence for the review text. This release contains none of it. - Questions, answers, label maps and documentation written for TR-OOLONG: CC-BY-4.0, except where a set's share-alike licence applies. The synthetic user IDs and dates are random and do not refer to anyone. ## Using the results - Report results per set, per question type and per length; do not pool them. Some question types can be answered well from a small sample of the records (most/least common label, A vs B at large margins); exact counts and user/date questions cannot. - A correct answer does not prove that a model read the whole document. - In the extension, report the 0% (tie) cells separately, report the two renderings separately, and compute intervals over documents (a document's two renderings, and MASSIVE's two languages, are one unit). - Labels-shown documents measure counting over the whole margin sweep. Labels-hidden results at 0.1% and 0.5% belong in a separate stress test (see below), not with the other margins. - For labels-hidden runs, give the model only `text` and `question`: no answers, counts, other documents or source lookup. The dataset is not protected against models that have seen these files or the sources. - The labels are the source's ratings, checked on samples; some are wrong. At the smallest margins, a few wrong source labels can decide the answer. ## Shortcut checks on the extension Each reader below was run on every test document, given the true labels of the records it reads (share of documents answered correctly; ranges over sets and record counts; 20 documents per cell, so about plus or minus 0.2 for a single reader). | reader | 0% (tie) | 0.1-0.5% | 2% | 10% | |---|---:|---:|---:|---:| | every record | 1.00 | 1.00 | 1.00 | 1.00 | | random 25% of records | 0.01-0.05 | 0.49-0.70 | 0.65-0.98 | 0.99-1.00 | | random 5% of records | 0.03-0.08 | 0.45-0.59 | 0.53-0.80 | 0.79-1.00 | | question text alone | 0.00 | 0.45-0.55 | 0.35-0.50 | 0.35-0.50 | | the label with more source records | 0.00 | 0.40-0.60 | 0.30-0.50 | 0.40-0.50 | | word-count classifier on every record, then exact count | 0.00 | 0.45-0.65 | 0.30-0.85 | 0.50-1.00 | | every record, 1 in 100 of one label misread | 0.00-0.01 | 0.50-0.71 | 1.00 | 1.00 | | every record, 5 in 100 of one label misread | 0.00 | 0.50 | 0.50-0.54 | 1.00 | Also checked: reading the first, last, evenly spaced or first-and-last part of a document scores on average within 0.03 of a random sample of the same size. The association between a record's label and its month, weekday, user frequency or line position is at most 0.02 (Cramér's V); the largest relative to reshuffled labels is MASSIVE's user frequency, V = 0.009, just above the reshuffled 95th percentile of 0.008. Counting mentions of a label's name scores 0.20-0.65. No record is in both the development and the test documents. Two limits follow. **Memory:** records are drawn from the same pools for every document, so nearly every test record also appears, with its label shown, in another released document (and every label is in the public source dataset). A reader that remembers those labels answers almost every question; the labels-hidden condition is meaningful only for models not trained on these files or the sources. **Label errors:** below about 2%, the answer depends on every record being read exactly as the source labels it. A reader that misreads even 1 in 100 records of one label in one direction scores 0.50-0.71 at 0.1-0.5%, close to a coin flip, while the source labels agree with a human reader on 88-100% of review records and 79-84% of MASSIVE records. Labels-hidden results at 0.1-0.5% therefore measure agreement with the source labels under exact counting, which is a stress test, not evidence of reliable classification. ## Changes from v0.10 v0.11 rebuilds the dataset along OOLONG-synth's design (random label mix per document, users and dates on every record, counting/user/timeline questions); it is not comparable question by question with v0.10. Sets were re-chosen after a blind label review: the complaint and news sets are removed, MASSIVE uses 11 scenario groups, sentiment is positive vs negative. MARC text is no longer distributed. ## Citation Cite this dataset (citation added with the paper) and the sources: OOLONG (Bertsch et al., 2025, arXiv:2511.02817), MASSIVE (FitzGerald et al., ACL 2023), the Turkish NLP Suite datasets (Altinok, ACL 2023), the Multilingual Amazon Reviews Corpus (Keung et al., EMNLP 2020), Amazon Reviews 2023 (Hou et al., 2024).