SBLGNT_to_FALAM

LoRA adapter for Bible translation: SBLGNT → FALAM

Model Details

• Base Model: google/madlad400-3b-mt
• Task: Machine Translation (English → Falam Chin)
• Training Data: Bible verse pairs
• Method: LoRA (Low-Rank Adaptation)

Training Details

• Training Pairs: 6,788
• Validation Pairs: 61
• Validation Strategy: SingleBookStrategy(book=2PE)
• Epochs: 20
• Learning Rate: 0.0003
• LoRA Rank: 32
• LoRA Alpha: 64

Metrics

Metric	Score
BLEU	19.24
chrF	48.07
Levenshtein Similarity	0.54

Usage

from transformers import T5ForConditionalGeneration, T5Tokenizer
from peft import PeftModel

# Load base model and tokenizer
model_name = "google/madlad400-3b-mt"
model = T5ForConditionalGeneration.from_pretrained(model_name)
tokenizer = T5Tokenizer.from_pretrained(model_name)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "fotcode/SBLGNT_to_FALAM")

# Translate
input_text = "<2cfm> In the beginning God created the heavens and the earth."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=128)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

Citation

If you use this model, please cite:

@misc{bible-translation-lora,
  title={Bible Translation LoRA Adapters},
  author={LRL Translation Project},
  year={2025},
  url={https://huggingface.co/fotcode/SBLGNT_to_FALAM}
}