d-michail/mitos-qa

Mitos QA

Greek question-answering data about administrative procedures of the Greek public administration, built from the National Registry of Administrative Procedures (ΕΜΔΔ, mitos.gov.gr; open data API: https://api.mitos.gov.gr/info), which is public data. The snapshot was downloaded on 1-2 October 2026 (4,407 procedures, last updated between 26 June and 1 October 2026). It has two training sets (chat-format dialogs, and template questions answered with the registry's own text) and an evaluation benchmark. Code and documentation: https://github.com/d-michail/mitos-qa-dataset.

License and attribution

Released under CC BY-SA 4.0 (license), the license of the registry content. Attribution: built from data of the Εθνικό Μητρώο Διοικητικών Διαδικασιών – Μίτος (mitos.gov.gr; open data API: https://api.mitos.gov.gr/info), content licensed under CC BY-SA 4.0; according to the terms of use of mitos.gov.gr the data belong to, or are supervised by, the Hellenic Ministry of Interior. Changes: the registry texts were split into topics and processed into dialogs and question-answer records (and benchmark questions). Data snapshot: downloaded on 1-2 October 2026; the records were last updated between 26 June and 1 October 2026. The registry does not guarantee completeness or correctness (terms of use, A.6): verify answers against the live procedure at https://id.mitos.gov.gr/<Mitos ID>.

Configurations

Config	What it is	Records
qa-train	citizen / clerk dialogs, held-out processes excluded	85,768
qa-all	the same dialogs for all processes (train + heldout splits)	90,145
json-train	template questions answered with the registry's text, held-out processes excluded	85,768
json-all	the same for all processes (train + heldout splits)	90,145
benchmark	evaluation questions with reference answers	1,933

*-train is the data of the benchmark models; *-all is the data of the final models (trained on everything). The train split of *-all is identical to the *-train data.

Held-out processes. 220 of the procedures were held out, chosen at random, stratified by the owning organization (seeded). Their ids are in holdout-processes.txt (SHA-256 f8b360db55dd7f21cb7ea7e3e76339420be377b595c8bc3a04169f7abe7265b5), and every record has a process_id and a split (train or heldout).

Warning: do not evaluate on the benchmark a model that was trained on the heldout records. The unseen questions of the benchmark are written from exactly those 220 procedures, so a model trained on qa-all or json-all has seen their facts. Use qa-train / json-train for models you want to benchmark. The benchmark is for evaluation only.

Data format

qa-* and json-* records (JSON lines):

{"id": "100201.process-rules.3", "process_id": "100201", "source_url": "https://id.mitos.gov.gr/100201", "generator": "process-rules", "split": "train",
 "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

id is {process_id}.{generator}.{index}; process_id is the 6-digit Mitos ID of the procedure and source_url its official page, so every record can be traced to the source; generator is the topic of the procedure the dialog is about (rules, conditions, evidences, costs, owner organization, and so on). messages alternate user / assistant, starting with user.

benchmark records: id, split (seen: questions about processes in the training data; unseen: about held-out processes; open: questions from gov.gr life events, without a reference answer), process_id, generator, question, reference, key_facts (short facts copied from the source text), reference_contexts (the source text), gold_chunk_ids, and for open questions life_event and sub_event. Split sizes: open 595, seen 898, unseen 440.

How it was made

1. All procedures were downloaded from the ΕΜΔΔ API (4,407 processes).
2. Each procedure was turned into topic-specific prompts (one generator per field of the registry).
3. qa-*: each prompt was given to Qwen/Qwen3.8-27B_H100 (temperature 0.7), which wrote one short citizen / clerk dialog in Greek grounded in the procedure's text; the dialogs were parsed into chat records, and malformed ones were regenerated.
4. json-*: one fixed Greek question template per topic (no language model), for example "which documents are needed for the procedure X?"; the answer is the registry's own text for that topic, the fields of the procedure rendered as structured text. The quick-info topic answers with the whole procedure record, so some answers are very long (up to tens of thousands of tokens).
5. benchmark: for a sample of the training processes (seen) and for all held-out processes (unseen), Qwen/Qwen3.8-27B_H100 wrote one question, a reference answer and key facts from a single topic chunk of the procedure; seen questions that were too similar (embedding similarity above 0.92) to a training question of the same process were rejected.

Limitations

• The dialogs, the benchmark questions and their references were all written by the same language model, so they inherit its mistakes and style; they were not checked by people, and the data may contain errors or outdated information. It is not legal advice.
• The registry changes over time; this is a snapshot.
• The same model family wrote the benchmark and may judge answers to it, which can favor models of that family; use more than one judge.
• The benchmark was written from the same procedure text that retrieval systems index, which suits RAG evaluation but means the questions are not independent of that text.