Property	Value
Base Model	google/gemma-3-4b-it
Translation Direction	English → isiZulu
LoRA Rank (r)	8
LoRA Alpha	16
Training Method	QLoRA (4-bit quantization)
Domain	Scientific/Academic texts

Why LoRA?

LoRA (Low-Rank Adaptation) enables efficient fine-tuning by training only a small number of additional parameters. This adapter adds only ~4.0M parameters to the base model while achieving strong translation performance.

Evaluation Results

Performance on the AfriScience-MT test set:

Split	BLEU	chrF	SSA-COMET
Validation	19.79	50.98	61.18
Test	20.21	52.61	60.67

Metrics explanation:

BLEU: Measures n-gram overlap with reference translations (0-100, higher is better)
chrF: Character-level F-score, robust for morphologically rich languages (0-100, higher is better)
SSA-COMET: Neural metric trained for Sub-Saharan African languages, shown as percentage (0-100, higher is better) (McGill-NLP/ssa-comet-stl)

Usage

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

# Configure 4-bit quantization (recommended for memory efficiency)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-4b-it",
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-4b-it")

# Load LoRA adapter
adapter_name = "dsfsi/gemma_3_4b_it-lora-r8-eng-zul"
model = PeftModel.from_pretrained(base_model, adapter_name)
model.eval()

# Prepare translation prompt
source_text = "Climate change significantly impacts agricultural productivity in sub-Saharan Africa."
instruction = "Translate the following English scientific text to isiZulu."

# Format for Gemma chat template
messages = [{"role": "user", "content": f"{instruction}\n\n{source_text}"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Generate translation
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        num_beams=5,
        early_stopping=True,
        pad_token_id=tokenizer.pad_token_id,
    )

# Decode only the generated part
generated = outputs[0][inputs["input_ids"].shape[1]:]
translation = tokenizer.decode(generated, skip_special_tokens=True)
print(translation)

Without Quantization (Full Precision)

python

# For GPUs with sufficient memory (>24GB for larger models)
base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-4b-it",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base_model, "dsfsi/gemma_3_4b_it-lora-r8-eng-zul")

Training Details

Hyperparameters

Parameter	Value
LoRA Rank (r)	8
LoRA Alpha	16
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Epochs	3
Batch Size	2
Learning Rate	2e-04
Max Sequence Length	512
Gradient Accumulation	4

Hardware Requirements

Configuration	VRAM Required
4-bit (QLoRA)	~8-12 GB
8-bit	~16-20 GB
Full precision	~24-40 GB

Reproducibility

To reproduce this adapter:

bash

# Clone the AfriScience-MT repository
git clone https://github.com/afriscience-mt/afriscience-mt.git
cd afriscience-mt

# Install dependencies
pip install -r requirements.txt

# Run LoRA training
python -m afriscience_mt.scripts.run_lora_training \
    --data_dir ./data \
    --source_lang eng \
    --target_lang zul \
    --model_name google/gemma-3-4b-it \
    --model_type gemma \
    --lora_rank 8 \
    --output_dir ./output \
    --num_epochs 3 \
    --batch_size 4 \
    --load_in_4bit

Limitations

Domain Specificity: Optimized for scientific/academic texts; may underperform on casual or colloquial language.
Language Direction: Only supports English → isiZulu translation.
Base Model Required: Must be used with the google/gemma-3-4b-it base model.
Context Length: Maximum context is model-dependent; longer texts should be chunked.

Citation

If you use this model, please cite the AfriScience-MT paper (arXiv:2605.29741):

bibtex

@article{abdulmumin2026afriscience,
  title   = {AfriScience-MT: Towards Decolonizing Science in Africa through Text Translation},
  author  = {Abdulmumin, Idris and Gwadabe, Tajuddeen and Muhammad, Shamsuddeen Hassan and Adelani, David Ifeoluwa and Khalo, Nomonde and Ahmad, Ibrahim Said and Modupe, Abiodun and Mumm, Anina and Biyela, Sibusiso and Rabie, Michelle and Havemann, Johanna and Rei, Marek and Abbott, Jade and Marivate, Vukosi},
  journal = {arXiv preprint arXiv:2605.29741},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.29741}
}

License

This adapter is released under the Apache 2.0 License.

Acknowledgments

Base model: google/gemma-3-4b-it
LoRA implementation: PEFT
Evaluation: SSA-COMET for African language assessment