Neural machine translation between Spanish and Inga, a Colombian Indigenous language (Inga people, Putumayo and Nariño). The model is bidirectional: it translates Spanish to Inga and Inga to Spanish.

Fine-tune of facebook/nllb-200-1.3B from the thesis Data-Centric Strategies for Low-Resource Neural Machine Translation in Colombian Indigenous Languages (Universidad de los Andes, FLAG Lab).

Languages

Role	Language	Tokenizer code	ISO 639-3
Source / target	Spanish	`spa_Latn`	spa
Source / target	Inga	`ing_Latn`	inb

Results

SacreBLEU and chrF2 on the held-out test partition.

Direction	BLEU	chrF2
Spanish to Inga	1.47	27.68
Inga to Spanish	4.62	23.73

Usage

python

from transformers import AutoModelForSeq2SeqLM, NllbTokenizer

tok = NllbTokenizer.from_pretrained("Flaglab/nllb-1.3B-inga-esp")
model = AutoModelForSeq2SeqLM.from_pretrained("Flaglab/nllb-1.3B-inga-esp")

# Spanish to Inga
tok.src_lang = "spa_Latn"
inputs = tok("Buenos días, ¿cómo estás?", return_tensors="pt")
out = model.generate(**inputs, forced_bos_token_id=tok.convert_tokens_to_ids("ing_Latn"))
print(tok.batch_decode(out, skip_special_tokens=True))

Training

Base model: facebook/nllb-200-1.3B
Strategy: Expanded bilingual corpus (full variant).
Tokenizer: NLLB SentencePiece extended with a custom Inga vocabulary.

License and intended use

Released under CC-BY-NC 4.0, inherited from NLLB-200: non-commercial use only. Intended for research on machine translation for Colombian Indigenous languages. Outputs have not been validated by human speakers.