aoiandroid/africa-v1-fused-model

Africa v1 Fused Model

This is the fully merged (fused) version of the Africa v1 translation model, combining the Qwen3-4B-Instruct-2507 base model with LoRA adapters.

Model Description

This model is a fused version of the Africa v1 translation model, where the LoRA adapters have been merged into the base model weights. This provides:

• Standalone model - No need to load separate adapters
• Full precision - Model weights in safetensors format (not quantized)
• Direct deployment - Ready to use with standard inference frameworks

The model supports translation between English and 29 African languages.

Format

• File Format: SafeTensors
• Size: ~2.1 GB
• Quantization: None (full precision from MLX 4-bit base)
• Architecture: Qwen3-4B with merged LoRA weights

Supported Languages (29)

African Languages:

• Afrikaans (af), Akan (ak), Amharic (am), Bambara (bm), Ewe (ee)
• Fula (ff), Hausa (ha), Igbo (ig), Kinyarwanda (rw), Kirundi (rn)
• Kongo (kg), Lingala (ln), Luganda (lg), Ndebele (nd), Northern Sotho (nso)
• Chichewa/Nyanja (ny), Oromo (om), Shona (sn), Somali (so), Swahili (sw)
• Tigrinya (ti), Tsonga (ts), Tswana (tn), Twi (tw), Venda (ve)
• Wolof (wo), Xhosa (xh), Yoruba (yo), Zulu (zu)

Plus English (en) for bidirectional translation.

Training Details

Base Model

• Model: Qwen3-4B-Instruct-2507
• Parameters: 4 billion
• Architecture: Transformer-based language model

LoRA Fine-tuning (Merged)

• LoRA Rank: 8
• LoRA Alpha: 20
• Target Layers: 16 layers
• Training Iterations: 10,000
• Learning Rate: 5e-5

Fusion Process

The model was created by:

1. Training LoRA adapters on the MLX 4-bit quantized base model
2. Merging adapters into base weights using mlx_lm.fuse
3. Exporting as standalone model with config and tokenizer

Usage

HuggingFace Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("aoiandroid/africa-v1-fused-model")
tokenizer = AutoTokenizer.from_pretrained("aoiandroid/africa-v1-fused-model")

# Prepare translation prompt
prompt = "Translate from English to Swahili:\n\nHello, how are you?"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate translation
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

MLX (Apple Silicon)

from mlx_lm import load, generate

# Load model
model, tokenizer = load("aoiandroid/africa-v1-fused-model")

# Generate translation
prompt = "Translate from English to Swahili:\n\nHello, how are you?"
response = generate(model, tokenizer, prompt=prompt, max_tokens=256, temp=0.1)
print(response)

vLLM (Fast Inference)

from vllm import LLM, SamplingParams

# Initialize model
llm = LLM(model="aoiandroid/africa-v1-fused-model")

# Generate translation
sampling_params = SamplingParams(temperature=0.1, max_tokens=256)
outputs = llm.generate(
    ["Translate from English to Swahili:\n\nHello, how are you?"],
    sampling_params
)

for output in outputs:
    print(output.outputs[0].text)

Advantages Over Other Formats

Format	Size	Standalone	Speed	Precision
Fused Model	2.1 GB	Yes	Fast	Full (from 4-bit base)
LoRA Adapters	29 MB	No	Fast	N/A
GGUF Q4_K_M	2.3 GB	Yes	Very Fast	4-bit
MLX 4-bit	2.1 GB	Yes	Very Fast	4-bit

Use this format when:

• You want a standalone model without separate adapters
• You're using standard inference frameworks (Transformers, vLLM)
• You need compatibility with cloud deployment services

Evaluation

Same evaluation results as Africa v1:

Metric	Score	Interpretation
Non-empty outputs	30/30 (100%)	All samples generate output
BLEU	0.71	Very low - experimental model
chrF	9.24	Low character-level overlap
TER	362.19	High edit distance

See Africa v1 model card for detailed evaluation.

Limitations and Biases

• Experimental Model: This is v1 with known quality issues
• Repetition: Model may get stuck in repetition loops
• Hallucination: May generate fluent but incorrect translations
• Low-Resource Languages: Limited training data for some African languages
• English-Centric: Best performance on English↔African pairs

Intended Use

• Research: Exploring multilingual translation for African languages
• Experimentation: Testing deployment with standard frameworks
• Development: Building translation applications (with quality caveats)

Not recommended for production use. Use v2 or specialized translation models for production.

Model Variants

This repository contains the fused model. For other formats:

• GGUF + LoRA + MLX: africa-v1-translation-model
• Improved v2: africa-v2-translation-model

Citation

@software{africa_v1_fused_model,
  title = {Africa v1 Fused Translation Model},
  author = {TranslateBlue Project},
  year = {2026},
  url = {https://huggingface.co/aoiandroid/africa-v1-fused-model}
}

License

Apache 2.0

Model Card Authors

TranslateBlue Project

Model Card Contact

For questions or issues, please open an issue in the model repository.