birgermoell/oellm-eu-tasks

OpenEuroLLM EU task suites

Two verifiable task suites in ten EU languages, built for evaluating and RL-training ~10B-parameter assistants. They follow the spirit of the Artificial Analysis Intelligence Index (knowledge with a penalty for guessing; practical document work), at a difficulty where small models are not at the floor. Every item has an answer that a program can check, so the same data serves as a held-out benchmark (test) and as an RL pool with a programmatic reward (train).

Config	What	Test	Train	Scoring
qa	Calibrated factual QA about European people, places, works and organisations (from Wikidata)	600	30,791	+1 correct, 0 abstain (ANSWER: UNKNOWN), −1 wrong
doc	Generated business documents: invoices, staff tables, contracts, room bookings, letters → JSON	500	6,000	exact answer 1/0; JSON: fraction of fields

Languages: en, de, fr, es, it, nl, sv, pl, pt, fi (test: 60 QA items and 50 document tasks per language).

The scoring code, the generators and the RL verifiers (eu_qa, eu_doc) are in oellm-rlvr/src/oellm_rlvr/eu_tasks.py (Apache-2.0); the eval runner is qwen35-posttrain/scripts/eval_eu_tasks.py. Please score with that code: the answer parsing (locale number formats, aliases, abstentions) is part of the benchmark.

qa: calibrated factual QA

Seven question types: birth year, birthplace, capital/administrative seat of a region, film director, author of a literary work, founding year (universities, football clubs, companies, museums), and where a river flows into. Subjects come from 17 European countries (SE, FI, DK, DE, FR, ES, IT, NL, PL, PT, AT, BE, CZ, HU, RO, GR, IE) in three popularity bands by the subject's Wikidata sitelinks: head ≥ 30, torso 10–29, tail 3–9. The tail is mostly unknowable for a small model; the right behaviour there is to abstain.

Each prompt asks for the name or year on a final line ANSWER: <answer> and says that ANSWER: UNKNOWN costs nothing while a wrong answer counts against the model. Years must match exactly (a reply with two years is wrong); names match any Wikidata label or alias of the answer entity in any of the fetched languages (case, accents and punctuation normalised; a short reply that contains the name, such as "Uppsala, Sweden", counts; a list of guesses does not). A reply without an ANSWER: line counts as wrong.

Construction: one SPARQL query per type × country × band (Wikidata Query Service, fetched 2026-10-27), labels and aliases from wbgetentities. A fact is kept only if it has exactly one answer (one birth year, one director, …), and dropped if any answer alias occurs in the subject's own name ("Uppsala County" → "Uppsala"). People are disambiguated by occupation and country, organisations by type and country, films by year and country. Subjects are split by hash: 15 % test candidates, 85 % train, so no subject appears in both. The test set is balanced over type × band × language (200 items per band).

Columns: id, question_type, language, popularity_band, subject_sitelinks, prompt, answer_kind (year | entity), answers (accepted strings), source (wikidata:Q…).

Report: accuracy, abstention rate, hallucination rate = wrong / (n − correct), and the knowledge index 100 × (correct − wrong) / n (0 = always abstain), overall and per band.

doc: document tasks

Procedurally generated documents with locale-formatted numbers and dates (1.234,56, 1 234,56, 14.03.2026, …) and a question with an exact answer:

Family	Question variants (* = held out of train)
invoice	total incl. VAT of all invoices; total VAT; invoice with the highest total; net total for one customer*
table (staff register)	count by department and city; salary sum of a department; earliest start date in a department; average salary of a department*; count started after a date
contract	end date of a contract (ISO); contracts ending before a date; contract that ends last*
meeting (room bookings)	the two overlapping meetings; booked minutes of a room on a day*
form	extract name, customer number, order date, amount and city from a letter as JSON (with a distractor order)

Levels 1–3 scale the number of invoices, rows, contracts and bookings. Answers: amounts with two decimals (±0.01), integers, ISO dates, ID sets, names, or JSON fields. The test split contains every variant (120 items of held-out variants); train contains only the others, so the test also measures transfer to an unseen question type. Test and train use disjoint generator seeds.

Columns: id, family, variant, held_out_variant, language, level, prompt, answer_kind, label (JSON with the gold value and, for IDs, the ID pattern).

Baselines (greedy; eval_eu_tasks.py, strict ANSWER: line)

Model	Mode	QA accuracy	QA head acc.	Abstain	Hallucination	QA index	Doc score
oellm-9b-256k-stage5-a4b-tool-rl-step48	default	0.265	0.500	3 %	96 %	−43.7	0.360
same	thinking	0.238	0.460	3 %	96 %	−49.0	0.289
OpenEuroLLM 9B W1a (internal RL checkpoint)	default	0.273	0.515	3 %	95 %	−42.0	0.378
same	thinking	0.270	0.505	5 %	93 %	−40.8	0.351
Qwen3-8B	default	0.157	0.310	15 %	82 %	−53.3	0.742
Qwen3-8B	thinking	0.205	0.425	24 %	70 %	−35.5	0.947

Sources and licences

• qa: facts, labels and aliases from Wikidata (CC0 1.0); questions, templates and splits by the dataset authors.
• doc: fully synthetic (names, companies, amounts and dates are random; any resemblance to real entities is coincidental).
• This dataset: CC BY 4.0. Code: Apache-2.0.

Known limitations: Wikidata can be wrong or incomplete (a correct but undocumented answer is scored wrong); entity matching is string-based; question templates were written by the authors and not reviewed by native speakers for every language; the strict ANSWER: keyword also measures instruction following (some models translate it).

Canary string (please do not train on the test splits): oellm-eu-tasks-canary 18d6e8c1-9f4d-462b-b1d2-ad06986324ab

Part of the OpenEuroLLM post-training work at AI Sweden.