nllb-600M-pms-avg

Zero-shot baseline: NLLB-200 (distilled, 600M) with a new pms_Latn language token. No training on Piedmontese.

Condition NLLB-avg in a study on Piedmontese MT and its evaluation.

Condition	Model
NLLB-avg	this model
NLLB-avg+MT	nllb-600M-pms-avg-mt
NLLB-avg+DAE-50K+MT	nllb-600M-pms-avg-dae50k-mt
NLLB-avg+DAE-full+MT	nllb-600M-pms-avg-daefull-mt

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

repo = "offelia39/nllb-600M-pms-avg"
tok = AutoTokenizer.from_pretrained(repo, src_lang="ita_Latn")
model = AutoModelForSeq2SeqLM.from_pretrained(repo)

inputs = tok("La superficie della Luna è costituita da rocce e polvere.", return_tensors="pt")
out = model.generate(
    **inputs,
    forced_bos_token_id=tok.convert_tokens_to_ids("pms_Latn"),
    num_beams=4,
    max_new_tokens=128,
)
print(tok.batch_decode(out, skip_special_tokens=True)[0])

pms_Latn is not one of NLLB-200's languages. It was added for this work, so load the tokenizer from this repo, not from the base model.

License

CC BY-NC 4.0, inherited from facebook/nllb-200-distilled-600M.

Code and citation

Code: https://github.com/panchenko-ling/piedmontese_wiki_ml

@inproceedings{panchenko2026metrics,
  author    = {Panchenko, Yulia},
  title     = {When metrics miss the language: {M}achine translation evaluation for {P}iedmontese},
  booktitle = {Proceedings of the Twelfth Italian Conference on Computational Linguistics},
  series    = {CLiC-it '26},
  address   = {Palermo, Italy},
  publisher = {CEUR Workshop Proceedings},
  year      = {2026},
  pubstate  = {forthcoming},
}

@mastersthesis{panchenko2026thesis,
  author = {Panchenko, Yulia},
  title  = {Low-Resource Machine Translation for {P}iedmontese: Model Training, Human Evaluation, and Metric Reliability},
  school = {University of Turin},
  year   = {2026}
}