k3dani/hu-general-decision-mix

Hungarian general decision mix

Typed decisions with an explicit "none of these" answer: three quarters of the items come from 21 Hungarian task families, one quarter from an English backbone of 186 small families. Each item is a state, a question and 2–10 options. The answer is one option, or X when none of them is correct.

This is the training and validation data of a pre-registered experiment that failed. We trained one decision adapter on it to see whether a general adapter brings task-specific gains to task families it never saw. It did not: on held-out groups its precision fell below our failure line, on two of them mainly because it confidently picked an option where the answer was "none of these". The adapter is published as a research artifact, k3dani/Qwen3.6-35B-A3B-hu-general-decision-lora-negative-result, not for use. The data itself has no known defect beyond those listed under Known issues. We release it so the result can be reproduced, challenged or improved on.

Magyarul röviden: magyar (és 25% angol) döntési itemek „egyik sem” opcióval, sok feladatcsaládból: egy negatív eredményű, előre rögzített kísérlet tanító- és validációs adata. Ügyféladat nincs benne, a licenc soronként jelölt.

Write-up

• Blog: The general decision adapter that didn't generalise (magyarul)
• Evidence: pre-registered protocol, code and aggregate results in docai-evals

At a glance

Two configurations, one per licence group. Load both for the full training set.

config	split	items	scored (non-ambiguous)	X gold	Hungarian / English
permissive	train	11,066	10,760	1,077	7,006 / 4,060
permissive	validation	1,908	1,854	151	1,230 / 678
sharealike	train	6,434	6,429	542	6,119 / 315
sharealike	validation	1,214	1,213	115	1,164 / 50

• train (17,500 items) is exactly what the adapter was trained on.
• validation (3,122 items) is the split its temperature and threshold were fitted on. It comes from the same task families, plus two (HuSST, HuCommitmentBank) that were dropped from training.
• Not included: the held-out test groups of the study.
  • The product-matching test is published in k3dani/hu-invoice-catalog-decisions.
  • The tool-selection test contains benchmark rows (BFCL, When2Call) that we do not redistribute.
  • The KIE/validation test (synthetic invoices) is not included; the generator is in the evidence package.
• Shape: 207 task families (21 Hungarian, 186 English), 13,254 choice and 4,246 noul items (yes / no / X). Option counts run from 2 to 10; 8,222 items have two options.

Fields

{
  "id": "…", "family": "nli.hurte", "group": "nli", "lang": "hu", "kind": "choice",
  "state": "…", "question": "…",
  "options": [{"id": "o1", "text": "…"}, {"id": "o2", "text": "…"}],
  "gold": "o1", "x_type": null, "ambiguous": false, "soft": null, "group_id": "…",
  "source": "HuRTE", "source_ref": "…", "source_license": "CC-BY-SA-4.0", "license": "CC-BY-SA-4.0",
  "provenance": {}
}

• gold: an option id, or null when the correct answer is X ("none of these"). X is not listed in options. Add it last when you build the prompt.
• x_type: for X-gold items, how the X arose.
  • gold-drop: the correct option was removed from the list, and a related, not equivalent candidate remains (synonym-filtered by two LLM judges);
  • natural: the source itself marks the item unanswerable, e.g. MILQA, or the source's "none" answer;
  • out-of-scope: the request belongs to another task, with a superficially related option list;
  • missing-precondition: an option would fit, but the state lacks a fact needed to decide.
• ambiguous and soft: the source gives a label distribution rather than one label, or marks the item as ambiguous (366 items, all from the English backbone). These items are not scored. soft holds the distribution; during training it was the target.
• group_id: split unit. Permutations, counterfactual pairs and all questions on one state share it.
• group: the family group: nli, szentiment (acceptability in train; validation also has HuSST sentiment), qa, jogi (legal), kozbeszerzes (procurement), rag, level (letters and triage), entitas (entities), konstrukcios (text-structure constructions) or angol_gerinc (English backbone).
• source: the source's short name; for the English backbone tasksource:<subset>, procedural-typed-decisions or Open-Jev, and szint03 for our synthetic families.
• license: the licence of the row as released here. source_license is the licence of the source.
• provenance: source identifiers for attribution, e.g. celex, njt_id, wikipedia_title, tasksource_subset, cpv_division, recipe, counterfactual.

The exact prompt used for training and evaluation is decision_prompt.py in the adapter repository. The display order is the stored order (permutation 0), and the X option goes last.

Sources and licences

source	licence	train / validation
NerKor	CC-BY-SA-4.0	1,478 / 200
MultiEURLEX (hu)	CC-BY-SA-4.0	1,474 / 200
Wikipedia (hu, 20231101), text-structure constructions	CC-BY-SA-3.0	1,087 / 152
HuRTE, HuCOLA, HuWNLI	CC-BY-SA-4.0	1,756 / 240
HuSST, HuCommitmentBank	CC-BY-SA-4.0	0 / 172
MILQA (train part only)	CC-BY-SA-4.0	324 / 200
HuRC	CC-BY-4.0	1,224 / 163
HuCoPA	BSD-2-Clause	256 / 36
Nemzeti Jogszabálytár (Hungarian legislation)	not protected by copyright (Hungarian Copyright Act 1. § (4))	1,538 / 210
TED notices (Hungarian)	reuse under Commission Decision 2011/833/EU, contact details removed	1,110 / 150
tasksource/tasksource-jev-typed-decisions, full config	per row, commercial-use rows only (CC-BY, CC-BY-SA, MIT, Apache-2.0, BSD, ODC-BY, CC0)	2,216 / 329
tasksource/procedural-typed-decisions	Apache-2.0	1,136 / 198
ZefanCai/Open-Jev, CC0 control configurations	CC0-1.0	1,023 / 201
synthetic families (this release)	CC-BY-4.0	2,878 / 671

• The sharealike config holds every CC-BY-SA row. Derivatives of it stay under the same licence.
• The permissive config holds everything else. CC-BY rows require attribution; use source and provenance.
• Licences outside the SPDX list: Hungarian legislation (not protected by copyright) and TED notices (Decision 2011/833/EU) appear as other in the metadata; the per-row license field names them.
• Excluded by design:
  • rows whose labels came from a closed model or were distilled from one (for example tasksource/synthetic-typed-decisions);
  • NC / NC-SA sources;
  • evaluation-only benchmarks (Belebele, SIB-200, BFCL);
  • the routing, product and KIE/validation task groups, which were held out for testing.

How it was made

• Human-labelled Hungarian sources were converted to the common item format with per-source adapters. X items come from the source's own "unanswerable" marks or from gold removal.
• Text-structure constructions compute the gold from the text in Python: next sentence, section title, law, chapter and title.
• Synthetic families cover letter triage, document type, contract clause, entity matching, administrative case type, RAG relevance and answer review. They were written by Apache-2.0 models (Muse-Glimmer-30B; Mistral Small 4 for RAG relevance) from family specifications. The intended label is the gold candidate, and an independent Apache-2.0 model of another family (Gemma 4 26B-A4B-it) relabelled every item. An item was kept only if the checker's top label matched the intended one with probability ≥ 0.6. The Qwen family, which the studied adapter is built on, neither generated nor checked any item.
• No teacher model and no closed-model or Jev outputs appear in any label or soft target.
• Gold-drop X items were synonym-filtered: every remaining option was judged by Mistral Small 4 and Gemma 4. Any option judged an acceptable answer was removed.
• Machine noise estimation ran per source. There was no human label audit.
• No customer data. Names, e-mail addresses and phone numbers in synthetic items are fictional, and e-mail domains use the reserved .example TLD. Any collision with real entities is accidental. Contact details were filtered from the TED notices. Rows taken from public sources keep whatever their source text contains.

Known issues

• Construction artefacts. Constructed X dominates: 962 of the 1,619 training X items are gold-drop. The natural and missing-precondition types are small (172 and 91).
• One generator and one checker. The synthetic families carry the style of that single generator family, and their labels rest on the checker's agreement.
• Uneven English backbone. Most en.ts.* families contribute only 13–14 items each.
• No option descriptions. The prompt format supports a use_when description per option, but no option in this release has one.

Citation

@misc{kiss2026hugeneraldecisions,
  title        = {Hungarian general decision mix: training data of a pre-registered negative result},
  author       = {Kiss, D{\'a}niel},
  organization = {K3Net Kft.},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/k3dani/hu-general-decision-mix}}
}

Please also credit the original sources listed above.