DziriNames — weights & production index

Arabic ↔ Latin transliteration of Maghrebi names, Algerian-first. This repo hosts the binary assets for the DziriNames project:

File	Size	Role
models/mag_ft_lr1e-4/best.pt	21 MB	default model — pure Maghrebi fine-tune, 79.5 % @10 on the 1,941-key Algerian benchmark
models/mag_mix50/best.pt	21 MB	fallback — 50 % world replay, holds world recall floors (72.9 % @10 world)
models/rebal_f2/best.pt	21 MB	pre-fine-tune baseline
data/nmt/name_index_dzns_v2_prod.pkl	69 MB	production index — 369,230 attested Arabic keys, serves 98.7 % of benchmark keys in 0.03 ms

The models are 5.55M-parameter seq2seq transformers (d_model 256, 3 layers, fp32, CPU-only, ~322 ms median per generated name). The system is retrieval-first: the index answers when it can, the model only handles unseen names.

Usage, benchmark, and full measurements: https://github.com/ElphenomenoCoding/DziriNames

@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}
}

Name data derive from openly licensed corpora: Algerian names and surnames dataset (Youcef Amoura, Kaggle, MIT) and Algerian Name Transcription Corpus (Taha Zerrouki, CC0-1.0), plus Wikidata-based world corpora (CC0).