This repository contains a LoRA adapter fine-tuned for English to Twi translation. It was trained from facebook/nllb-200-distilled-600M using Ghana NLP's ENGLISH_TWI_PARALLEL_TEXT dataset.

Project Collaboration

This model was built and first published by McLanor Jeff under the personal repository mclanorjeff/english-twi-nllb-lora.

The shared collaboration version for the project team is available under the Lanor-and-Nick organization:

Lanor-and-Nick/english-twi-nllb-lora

That organization repository is used for collaborative access and team-facing updates while this personal repository remains the original author copy.

Model Details

Task: English to Twi translation
Base model: facebook/nllb-200-distilled-600M
Adapter type: LoRA / PEFT
Source language tag: eng_Latn
Target language tag: twi_Latn
Training hardware: NVIDIA RTX 2060, 6 GB VRAM
Trainable parameters: 2,359,296 of 617,433,088 total parameters

Dataset

The training data comes from Ghana NLP:

Dataset: ENGLISH_TWI_PARALLEL_TEXT
Training examples: 5,475
Validation examples: 608
Source column: English text
Target/reference column: Twi translation

The data was cleaned, encoding issues were repaired where possible, and a validation split was created before training.

Evaluation

Best validation checkpoint by BLEU: checkpoint-1000

Metric	Value	Interpretation
BLEU	26.9278	Decent/useful translation quality for a low-resource pair.
chrF	50.5380	Strong useful character-level similarity for Twi spelling/morphology.
Eval loss	1.4805	Lower is better; validation error improved during training.

Translation does not use ordinary classification accuracy because several Twi translations can be valid for the same English sentence.

Sample Outputs

English input	Twi output
`Where are you?`	`Ɛhe na wowɔ?`
`For this reason, many people are interested in it.`	`Esiane eyi nti, nnipa pii ani gye ho.`
`The students learn very well because they want to improve their future lives.`	`Asuafoɔ no sua adeɛ yie ɛfiri sɛ wɔpɛ sɛ wɔn daakye asetena tu mpɔn.`

Usage

python

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

model_id = "mclanorjeff/english-twi-nllb-lora"
source_lang = "eng_Latn"
target_lang = "twi_Latn"

config = PeftConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id, src_lang=source_lang, tgt_lang=target_lang)
base_model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(base_model, model_id)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

text = "Where are you?"
inputs = tokenizer(text, return_tensors="pt").to(device)
forced_bos_token_id = tokenizer.convert_tokens_to_ids(target_lang)

with torch.no_grad():
    output = model.generate(
        **inputs,
        forced_bos_token_id=forced_bos_token_id,
        max_new_tokens=192,
        num_beams=5,
        no_repeat_ngram_size=3,
    )

print(tokenizer.batch_decode(output, skip_special_tokens=True)[0])

Limitations

This model is useful for demos, learning support, and experimentation, but it is not a perfect translator. It may struggle with idioms, informal English, names, emotional nuance, rare Twi phrasing, or contexts far outside the Ghana NLP training data. Human review is recommended for high-stakes translation.