Ugiat/madlad400-3b-mt-es_ES

Translation • 3B • Updated Oct 7 • 13

Note Start here. MADLAD-400-3B-MT fine-tuned to translate into Castilian Spanish (vosotros, móvil, ordenador, coche, zumo…) with the target tag <2es_ES>. v4 on main (recommended); v1 (most systematic Castilian choices) and v3 on branches. Use transformers 4.x.

Ugiat/madlad400-3b-mt-es_ES-ct2

Translation • Updated Oct 7 • 12

Note The same model converted to CTranslate2 int8 (2.9 GB): faster and lighter, for production translation on GPU or CPU.

Note 1.18M translation pairs from 8 languages (en, fr, pt, ca, it, de, gl, eu) into Castilian, built by back-translation: the Castilian side is real text from Spain. The training data of the model.

Note 18.7M clean sentences of Spanish as written in Spain (438M words), selected from MADLAD-400's Spanish split, each tagged with the Castilian features it contains.

Note Spain vs Latin America scores, dialect markers, languages and website scores for all 250.9M documents of MADLAD-400's Spanish split; castilian_selected marks the 70.1M documents used for the sentences.

Ugiat/spanish-variety-fasttext

Text Classification • Updated Oct 6 • 57

Note Fast document classifier: Spain (es_ES) vs Latin America (es_LATAM). It produced the scores in the annotations dataset. For documents, not single sentences.

google/madlad400-3b-mt

Translation • 3B • Updated Nov 27, 2023 • 148k • 245

Note Upstream: the original multilingual translation model this work fine-tunes (Google).

Note The sister model for Latin American Spanish (es-419).

Note Upstream: the web corpus all the Castilian data was selected from (Google, distributed by AllenAI).

Note Upstream: the MADLAD-400 paper (Kudugunta et al., 2023), which introduced both the corpus and the base model.