nllb-ewe-en-fr-multilingual-lora

nllb-ewe-en-fr-multilingual-lora — Ewe ⇄ English ⇄ French (NLLB LoRA)

A LoRA adapter for facebook/nllb-200-distilled-600M fine-tuned for multi-directional translation between Ewe (ewe_Latn), English (eng_Latn) and French (fra_Latn). A single adapter handles all six directions (Ewe as both source and target).

Model details

• Developed by: Romaric Nadjire — academic project (M1 AI, INF2229)
• Type: PEFT / LoRA adapter for an NLLB-200 seq2seq translation model
• Languages: Ewe (ee), English (en), French (fr)
• Fine-tuned from: facebook/nllb-200-distilled-600M
• License: CC-BY-NC-4.0 (inherited from the base model — non-commercial, academic use)

Evaluation

Test-set BLEU / chrF++ (sacreBLEU), per direction:

Ewe↔English is usable; Ewe↔French is weak (much less parallel French↔Ewe data) and should be considered experimental.

How to get started

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from peft import PeftModel

BASE = "facebook/nllb-200-distilled-600M"
ADAPTER = "romaricnadjire/nllb-ewe-en-fr-multilingual-lora"
CODE = {"ee": "ewe_Latn", "en": "eng_Latn", "fr": "fra_Latn"}

tokenizer = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForSeq2SeqLM.from_pretrained(BASE)
model = PeftModel.from_pretrained(model, ADAPTER).eval()

def translate(text, src, tgt, num_beams=4):
    tokenizer.src_lang = CODE[src]
    inputs = tokenizer(text, return_tensors="pt")
    out = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(CODE[tgt]),
        max_new_tokens=128, num_beams=num_beams,
    )
    return tokenizer.batch_decode(out, skip_special_tokens=True)[0]

print(translate("Ŋdi na mi", "ee", "fr"))

Training

• Base: facebook/nllb-200-distilled-600M (frozen) + LoRA adapter.
• Data: trilingual parallel corpus (Ewe / English / French), deduplicated with a canonical order-independent key and filtered (foreign-script, length ratio, copies, max length); held-out validation/test excluded from training (anti-leak).
• Trained on Kaggle GPU.

Data sources & attribution

• OPUS bible-uedin (ee↔en, ee↔fr) — public domain / OPUS terms.
• GhanaNLP navigation corpus (ee↔en) — CC-BY 4.0.
• Kaggle tchaye59 Ewe–English pairs (mixed sources incl. religious texts) — research use.
• English↔French pivot data and public-domain UDHR (glk360).
• Base model NLLB-200 (facebook/nllb-200-distilled-600M) by Meta AI — CC-BY-NC 4.0.
• masakhane/mafand (CC-BY-NC) was used for evaluation only, never for training.

License & citation

This adapter is a derivative of facebook/nllb-200-distilled-600M and is released under CC-BY-NC-4.0 (non-commercial). Please credit the base model (Meta NLLB-200) and the data sources above.

Academic / portfolio project. Not legal advice; verify licenses before any non-research or commercial use.