yacinebenak/dzirinames-benchmark

DziriNames benchmark — Arabic → Latin transliteration of Algerian names

1,941 Arabic name keys, 3,480 accepted Latin references from real Algerian civil-status data. The first public benchmark for a problem screening systems get quietly wrong: Algerian names are romanized by French-era civil-status conventions (عبد القادر → Abdelkader), not Mashreq conventions (Abdul Al-Qadir) — and models trained on world data fail on them.

Format

One JSON object per line (eval_dz.jsonl):

{"ar": "آدم نسيم", "refs": ["adem nassim"], "n_refs": 1,
 "refs_mots": [["adem"], ["nassim", "nassime", "noussim"]],
 "src": ["dzns_eval_large"], "s2": false, "type": ["prenom"],
 "fuite_base": false, "note": ""}

• ar — normalized Arabic key
• refs — accepted Latin spellings (1.79 per key on average, max 38)
• refs_mots — per-word attested spelling groups for the 595 compound names (enables composed scoring)
• type — prenom / patronyme
• fuite_base — true for the 231 keys that also occur in common Wikidata-derived training corpora (lets you measure leak-free: 1,710 keys)
• s2 — historic hard-stratum flag (190 keys)

Provenance & quality

Every reference is attested — drawn exclusively from two openly licensed corpora of real Algerian names, nothing generated by rule or model. 295 disputed couples were hand-adjudicated; the measured reference error rate of the source file is 0.20 %. The build is deterministic (fixed seed, byte-identical rebuilds). Full lineage: PROVENANCE.md, adjudications: verdicts.json.

These are registers of onomastic forms only — no birth dates, no addresses, no identifying attributes of any person.

Baseline results (Arabic→Latin, recall @10, composed references)

System	@10
DziriNames mag_ft_lr1e-4 (5.5M params, Maghrebi fine-tune)	79.5 %
DziriNames mag_mix50 (50 % world replay)	76.1 %
rebal_f2 (no Maghrebi fine-tune)	68.5 %
ByT5-small (300M), fine-tuned on the same world corpus	loses by 33 pts @20 vs the 5.5M model

Code, models and evaluation harness: https://github.com/ElphenomenoCoding/DziriNames

Sources & attribution

Algerian names and surnames dataset — Youcef Amoura (rykerdz), Kaggle, 2024, MIT. https://www.kaggle.com/datasets/rykerdz/algerian-names-and-surnames-dataset

Algerian Name Transcription Corpus — Taha Zerrouki (linuxscout), CC0-1.0. https://github.com/linuxscout/algerian-name-transcription-corpus

Citation

@misc{benakmoume2026dzirinames,
  author = {Benakmoume, Yacine},
  title  = {DziriNames: a benchmark and retrieval-first system for Arabic--Latin transliteration of Maghrebi names},
  year   = {2026},
  url    = {https://github.com/ElphenomenoCoding/DziriNames}
}