NLLB-200 Fine-tuned: Spanish + English -> Kichwa
Ajuste fino (fine-tuning) de facebook/nllb-200-distilled-600M para traduccion Espanol -> Kichwa e Ingles -> Kichwa, sobre un corpus trilingue de 4 365 pares.
Metricas de evaluacion (test set 873 pares)
| Direccion | BLEU | chrF2++ | ED-similitud |
|---|---|---|---|
| ES -> KI | 19.78 | 46.21 | 0.5815 |
| EN -> KI | 19.69 | 44.87 | 0.5641 |
Uso rapido
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
REPO = "Dasskp/nllb-kichwa-es-en"
tokenizer = NllbTokenizer.from_pretrained(REPO)
model = AutoModelForSeq2SeqLM.from_pretrained(REPO)
def translate(text, src_lang, tgt_lang="ki_Latn"):
tokenizer.src_lang = src_lang
inputs = tokenizer(text, return_tensors="pt")
tgt_id = tokenizer.convert_tokens_to_ids(tgt_lang)
output = model.generate(**inputs, forced_bos_token_id=tgt_id)
return tokenizer.decode(output[0], skip_special_tokens=True)
print(translate("Buenos dias, hermano.", src_lang="spa_Latn"))
print(translate("Good morning, brother.", src_lang="eng_Latn"))