aoiandroid/africa-v2-translation-model

Africa v2 Translation Model

This is an improved fine-tuned translation model for 29 African languages, based on Qwen3-4B-Instruct-2507 with enhanced training data.

Model Description

Africa v2 is an improved version of the Africa v1 translation model. Key improvements include:

• System prompts in training data to enforce direct translation behavior
• Regenerated training dataset with better formatting
• Available in MLX 4-bit format and LoRA adapters

Note: Training was interrupted at 1,000 iterations due to GPU OOM. The model represents a partial training checkpoint.

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 (MLX 4-bit quantized)
• Parameters: 4 billion
• Architecture: Transformer-based language model

Fine-tuning

• Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 8
• LoRA Alpha: 20
• Target Layers: 16 layers
• Training Iterations: 1,000 (interrupted, intended 10,000)
• Learning Rate: 5e-5
• Batch Size: 1
• Final Train Loss: 2.116
• Final Val Loss: 2.512

Training Data

• Total Translation Pairs: 283,986
• Format: Enhanced with system prompts
• System Message: "You are a translation assistant. Output only the translation without explanation."

Available Checkpoints

• Checkpoint at 1,000 iterations (latest)

Improvements Over v1

1. System Prompts: Training data includes system messages to suppress thinking mode
2. Better Formatting: Consistent prompt format for all language pairs
3. Direct Translation: Model trained to output translation directly without explanation

Evaluation Results

Status: Model evaluation pending. Use v1 evaluation as baseline comparison.

Expected improvements over v1:

• Reduced repetition loops
• Less hallucination due to system prompts
• More consistent output format

Usage

MLX with mlx-lm

# Install MLX
pip install mlx-lm

# Download model
huggingface-cli download aoiandroid/africa-v2-translation-model --local-dir africa-v2-mlx --include "mlx-4bit/*"

# Run inference with system prompt
python -m mlx_lm.generate \
  --model africa-v2-mlx/mlx-4bit \
  --prompt "<|im_start|>system
You are a translation assistant. Output only the translation without explanation.<|im_end|>
<|im_start|>user
Translate from English to Swahili:

Hello, how are you?<|im_end|>
<|im_start|>assistant" \
  --max-tokens 256 \
  --temp 0.1

LoRA Adapters

from mlx_lm import load, generate

# Load model with LoRA adapters
model, tokenizer = load("aoiandroid/africa-v2-translation-model", adapter_path="lora")

# Prepare prompt with system message
messages = [
    {"role": "system", "content": "You are a translation assistant. Output only the translation without explanation."},
    {"role": "user", "content": "Translate from English to Swahili:\n\nHello, how are you?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Generate translation
response = generate(model, tokenizer, prompt=prompt, max_tokens=256, temp=0.1)
print(response)

GGUF Conversion

Note: GGUF export is not available for this model because:

• mlx_lm.fuse --export-gguf only supports model_type in ["llama", "mixtral", "mistral"]
• Qwen3 has model_type: "qwen3", which is not yet supported

To convert to GGUF:

1. Export MLX model to HuggingFace format (if supported)
2. Use llama.cpp's convert_hf_to_gguf.py script
3. Or wait for mlx_lm to add Qwen3 support

Alternatively, use v1's GGUF model as a fallback.

Limitations and Biases

• Partial Training: Only 1,000 iterations completed (10% of planned training)
• Needs Evaluation: Translation quality not yet formally evaluated
• Low-Resource Languages: Limited training data for some African languages
• Experimental Model: Intended for research and experimentation

Intended Use

• Research: Studying impact of system prompts on translation quality
• Experimentation: Testing improved training data formatting
• Comparison: Baseline for comparing with fully trained models

Recommended: Complete training to 10,000+ iterations before production use.

Future Work

• Complete training to 10,000+ iterations
• Increase LoRA rank to 16 or 32
• Formal evaluation with BLEU, chrF, and TER metrics
• Compare performance against v1

Citation

@software{africa_v2_translation_model,
  title = {Africa v2 Translation Model},
  author = {TranslateBlue Project},
  year = {2026},
  url = {https://huggingface.co/aoiandroid/africa-v2-translation-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.