AT-Benefits

Open synthetic dataset for Austrian social-benefits navigation (SV KI-Lotse research lane): situations, multi-turn dialogues, documents, forms and legal QA in German (incl. Austrian dialect and plain language), English, Turkish, Bosnian/Croatian/Serbian, Romanian, Hungarian, Polish, Ukrainian and Arabic. Labels (benefits, offices, documents, citations) are computed by the SV KI-Lotse product engine, never by a language model; language models only write surface text, which is checked for language, faithfulness, leakage and near-duplicates. No real person is described. Labels are navigation hints, not eligibility decisions.

Config dialogues

Multi-turn, tool-exact pilot dialogues for Austrian social-benefits navigation. Every dialogue is an OpenAI-style messages list (system, user, assistant with tool_calls, tool) with the real brain tool contracts (lotse_*) and outputs from the research tool server, ending in a submit_plan call. Labels (benefits, offices, documents, citations, reward) come from the product engine and the gym reward, never from a language model.

• Tier O (oracle): the gym oracle's own trajectories (template questions and summaries in German, English or Turkish, exact tool calls and outputs, reward 1.0); no model text.
• Tier A (verbalized oracle): gym oracle trajectories; tool calls, tool outputs and the plan are unchanged, the bulk model wrote only the natural-language turns in the target language (German incl. Austrian dialect and plain language, English, Turkish, BCS, Romanian, Hungarian, Polish, Ukrainian, Arabic), with injected medical questions handed off to a doctor/1450 in a sample (safety_injected). In language_variant rows the user speaks a language other than the persona's; tool arguments keep the original situation text.
• Tier B (teacher): the strong model played the pilot in the gym against an LLM-simulated user with the real tools; kept only with gym reward >= 0.9, zero boundary flags and valid citations.
• Quality: every dialogue re-scored with the gym reward (reward, reward_breakdown); product boundary patterns (boundary_flags = 0 for all kept rows); language ID on every turn (turn_languages, language_match); MinHash near-duplicates marked (near_duplicate_of); readability of the pilot's turns; LLM judge rubric on every tier A dialogue (judge, only passing rows are released).
• Splits follow the persona splits; heldout_family holds the held-out benefit family.

Rows: 10352 ({'A': 351, 'B': 1, 'O': 10000}); tier A/B languages: {'de': 191, 'ar': 17, 'en': 38, 'pl': 16, 'tr': 23, 'bcs': 12, 'hu': 16, 'uk': 20, 'ro': 19}; mean reward of tier A/B 0.9975. Every tier A row was judged and passed the rubric (naturalness, faithfulness to tool outputs, language correctness, boundary compliance, plain language where requested; each >= 4 of 5); pass rate over all judged attempts 0.7374 (476 judged).

Release v0.2.

Config situations

1,000,000 synthetic Austrian household personas with a structured card, an opening situation text and engine-computed gold (benefits, offices, documents, minimum questions, criterion citations). Generator: the seeded SV KI-Lotse persona generator (seed 202611071), split by index before sampling; heldout_family holds every persona whose bundle contains area D (care), which never occurs in the other splits; test contains two phrasing styles unseen in training.

• text_source = "template": template text in German (incl. Austrian dialect), English or Turkish.
• text_source = "llm" (425 rows): the text was rewritten by gpt-6-luna in one of German (Austrian standard, Austrian dialect by state, plain language B1), English, Turkish, Bosnian/Croatian/Serbian, Romanian, Hungarian, Polish, Ukrainian or Arabic, matching the speaker's register, literacy and noise, from the persona card only. Gold for these rows was recomputed by the engine with the new text and language. Every LLM text passed language ID (lingua), no benefit/office names (gold + whole life-situation bundle), numbers only from the card, no contact data, length 8-220 words and MinHash near-duplicate removal, and was then judged by gpt-6-sol on faithfulness, no leakage, language, register, naturalness and concern (1-5 each); a row is kept only if faithfulness, no leakage and language score at least 4. Judge pass rate 0.8799 (95 % Wilson (0.8479, 0.9059), n=483); per-variant rates are in the metrics. Acceptance rate 0.85. Per-row quality fields are in quality (JSON).

Rows per split and text source: heldout_family/llm 26, heldout_family/template 18,156, test/llm 35, test/template 36,329, train/llm 333, train/template 908,757, validation/llm 31, validation/template 36,333.

Labels are navigation hints computed by product code, not eligibility decisions. Metrics, leakage test (text classifier vs card classifier, micro-F1) and a stratified spot-check sample are in situations/metrics.json.

Columns: id, split, goal, scenario, life_situation, style, noise (JSON), situation_text, template_text, text_source, text_language, text_variant, household (JSON), gold (JSON), gold_benefits, quality (JSON or null), seed.

Release v0.2.

forms

Filled official Austrian forms with field-level gold and field-detection gold.

• Templates. 27 registered templates of the SV KI-Lotse form registry: the seven SV-Box AcroForm applications (catalog schemas from packages/knowledge/forms) and the twenty flat SVS forms of the E13 gold set (hand-drawn answer areas, one-based pages, PDF points, top-left CropBox origin).
• Values come from the product. For each (template, persona) the confirmed facts of a synthetic AT-Benefits persona (with statement or document provenance and quotes) are passed to the product's prefill policy (proposeFields) and renderer (renderFormPdf) of packages/brain through a bun bridge; no value is chosen by a language model. The same bridge reproduces the product's E7 gold exactly (218/218 values, 1030/1030 blanks).
• Field gold (fields, JSON): field id, label, type, owner section, page, boxes, validation, required flag, signature flag, fact type and mapping kind (catalog, candidate for the dataset's conservative E13 mapping of the applicant's identity fields, unmapped), status (filled, missing:<reason>, provider_<owner>, signature), value and provenance. Doctor, hospital and office sections and signatures are never filled; IBANs only from the person's own statement.
• Detection gold: boxes, labels, types and owners for the flattened filled PDF (detection_flat) and for simulated scans (detection_scan: rotated, noisy, JPEG-compressed image PDFs; boxes transformed accordingly).
• PDFs: pdf_filled (product output, AcroForm), pdf_flat, pdf_scan — included only for forms whose publisher permits reuse with attribution (SVS: "Die Inhalte der SVS-Website dürfen im hier genannten Rahmen bis auf Widerruf weiterverwendet werden, um Quellenangabe wird jedoch ersucht"). SV-Box forms (ÖGK, PVA, FSW, AUVA) ship field gold only.
• Splits follow the persona splits (train, val, test, heldout_family).

Counts: 15120 instances, 4924 with PDFs, 6084 detection records. All persons are synthetic; numbers carry valid check digits but belong to nobody. Source attribution per row.

Release v0.2.

legal_qa

Graded multilingual retrieval questions over Austrian social-insurance law. Each row is a query whose answer is one exact RIS section (and Absatz where possible) of the corpus JulianAT/at-sv-law, with BM25 hard negatives.

• Generation. For each eligible section (2668 of 2,863; repealed and stub sections dropped) a seeded plan assigns 1-4 slots (language, style, difficulty easy/medium/hard); a language model writes the question, a German rendering, a short answer, the Absatz numbers and a verbatim evidence quote. Languages: German majority (standard, plain language, colloquial Austrian), English, Turkish, BCS, Romanian, Hungarian, Polish, Ukrainian, Arabic.
• Filtering (after arXiv 2601.14160): rule checks on every question (language ID, at most 6 consecutive tokens copied from the section, no § / Absatz / law-name citation, evidence quote found in the section, valid Absatz numbers, length), MinHash near-duplicate removal per language, and an LLM judge on every question (answerable only from the section, correct, self-contained, natural, no verbatim copying); only passing questions are kept: judge pass rate 0.9926 (n = 407, 95 % CI (0.9786, 0.9975)).
• Negatives. negative_ids: top-10 BM25 sections for the German rendering, excluding the positive and parallel provisions with the same title in other acts (parallel_ids, which may also answer the question). bm25_rank_de is the positive's BM25 rank (difficulty signal).
• Splits by section (no section in two splits): train/val/test = 90/5/5.

Counts: 404 questions (de 138, de-AT-dialect 40, de-easy 37, bcs 31, tr 30, ar 29, en 27, uk 20, ro 19, hu 17, pl 16). Law texts: Datenquelle: Rechtsinformationssystem des Bundes (RIS), Bundeskanzleramt Österreich, licence CC BY 4.0.

Release v0.2.

Config documents

Synthetic German (Austrian) documents a person brings to an advisor: hospital discharge letters (Entlassungsbriefe), decisions (Bescheide) with appeal instructions, requests for documents, reminders, private-practice invoices (Wahlarzt/Wahltherapeut Honorarnoten) and prescriptions. Each document is a real PDF (A4, standard fonts) with exact gold for documented-statement extraction (typed verbatim quotes with page, text offsets and bounding boxes) and plain-words letter explanation targets; about a third also come as degraded 150 DPI scans with OCR.

449 documents (heldout_family 35, test 22, train 368, val 24), 499 pages, 4296 gold statements.

Document type	Count
decision	138
discharge_letter	53
document_request	144
prescription	32
reminder	60
wahlarzt_invoice	22

How it is made

• Spec first. For each AT-Benefits persona (config personas; the split is the persona's split, so heldout_family documents concern held-out care benefits only) a seeded spec fixes every administrative fact: issuing office from the engine gold (office.resolve) with its catalog address, benefit, dates, case numbers, care level, amounts, deadlines, requested documents (catalog documents not held by the persona), medicine names, and the documented recommendations (benefits from the persona's gold, e.g. rehabilitation, care allowance, aids).
• Model writes prose only. gpt-6-luna writes the sections as JSON sentences with a role (clinical, recommendation, open task, administrative, other). Deterministic validation rejects output with dates or amounts not in the spec, medicine names in prose, recommendations that carry a medical rationale or are missing, or real public figures; rejected output is retried with the error list. Prescriptions are template-only.
• Renderer owns the gold. A deterministic reportlab layout (letterheads, address window, reference block, tables, footers, page breaks) draws every text item at a known position. Gold quotes, page-local [start, end) offsets and boxes come from the renderer, and every document is verified against the PDF text layer extracted with pdfium before release.
• Judge filter. gpt-6-sol scores every model-written document on realism, consistency, boundary, german, gold (1-5); only documents with all scores >= 4 are released. Pass rate so far: 417/549 (0.7596; Wilson 95 % [0.7221, 0.7934]).
• Near-duplicates of model-written text (MinHash, char 5-grams, Jaccard 0.85) are dropped.

Gold format (E12-compatible)

statements (JSON list): id, type (documented_recommendation | administrative_fact), fact_type, quote (verbatim page-text slice; line breaks inside a quote are \n), page (1-based), position {start, end} (page-local, Unicode code points = UTF-16 units here), bbox {x, y, width, height} normalised to the page with a top-left origin, per-line boxes, and provenance (source: spec, catalog or model). Fact types: institution 506, case_number 449, document_date 449, insurance_number 440, open_task 440, deadline 342, requested_document 294, named_medicine 219, next_step 204, decision 195, appeal_info 138, carrier 107, visit_date 103, documented_recommendation 98, amount 96, admission_date 53. pages follows the SV KI-Lotse brain's documentPageSchema (text = lines joined by newlines; rows are items within 5 pt of a baseline, left to right, like the E12 pdfjs serialization). exclusions lists clinical sentences (diagnoses, findings, course) for negative overlap checks: diagnoses and findings are never gold facts (PRD FR-D3 and the medical boundary). explanation holds FR-D8 targets computed from the spec, not by a model: what is asked (with catalog document ids), by when (absolute date or relative appeal period) and what happens next, plus a plain-language summary; for discharge letters the navigation targets (benefit and responsible office per documented recommendation).

Other columns: pdf (bytes), scan_pages (PNG bytes per page, grayscale 150 DPI), scan_meta (rotation, blur, noise, JPEG quality and the rotated gold boxes), ocr (Tesseract deu text per page, mean confidence, gold-quote recall), spec, sentences (model sentences with roles), judge, persona and benefit ids, token usage.

Limitations

• Synthetic. Hospitals, practices, rehabilitation centres and all persons are invented; carrier and authority letters name the real responsible office from the catalog but are not issued by it (footer on every page) and use no logos. Insurance numbers have a deliberately wrong check digit.
• Medical content is fictional and was not reviewed by clinicians; it is a distractor for extraction, not medical information. Appeal periods and courts follow Austrian practice as modelled for this dataset and are not legal advice.
• Layouts are generated, so real-world scan artefacts (stamps, handwriting, folds) are missing.
• Reproduce: uv run python -m datagen.documents.run --tranche 1 then uv run python -m datagen.documents.release (from ml/).

Release v0.2.