Sunbird / Sunflower-32B

🌻 Sunflower-32B Model Card

Model Description

🌻 Sunflower-32B is a multilingual language model developed by Sunbird AI for Ugandan languages. Built on the Qwen 3-32B architecture, the model supports translation and text generation across 31 Ugandan languages plus English. The model achieves the highest translation accuracy among evaluated models in 24 of 31 language pairs.

Developed by: Sunbird AI
Model type: Causal language model
Languages: 31 Ugandan languages + English (see language codes above)

Intended Uses

Primary Use Cases

• Translation between English and Ugandan languages
• Translation between Ugandan languages
• Text generation in Ugandan languages
• Question answering in Ugandan languages

Example Usage

import transformers
import torch

MODEL_PATH = 'Sunbird/Sunflower-32B'
SYSTEM_MESSAGE = 'You are Sunflower, a multilingual assistant made by Sunbird AI who understands all Ugandan languages. You specialise in accurate translations, explanations, summaries and other cross-lingual tasks.'

tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH)
model = transformers.AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map='auto',
)

instruction = "Translate from Luganda to English: Wano webawaaba?"

messages = [
    {"role": "system", "content": SYSTEM_MESSAGE},
    {"role": "user", "content": instruction}
]

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

inputs = tokenizer([prompt], return_tensors="pt").to('cuda')
outputs = model.generate(
    **inputs,
    max_new_tokens=500,
    num_beams=5,
    do_sample=True,
    temperature=0.5,
)

response = tokenizer.decode(outputs[0][len(inputs['input_ids'][0]):], skip_special_tokens=True)
print(response)

API Access

For production use, the model is available through the Sunbird AI API at https://api.sunbird.ai/

Training Details

Training Data

The model was trained on approximately 750 million characters of text collected from:

• Digitized books and educational materials
• Radio transcripts (500+ hours transcribed)
• Web data from MADLAD-400 and Common Crawl
• Existing multilingual datasets (SALT, FLORES-200, MT560, TICO-19)
• Dictionaries, proverbs, and cultural documents

Training Procedure

Stage 1: Continued Pretraining

• Base model: Qwen 3-32B
• Training time: ~6 hours on 4× H200 GPUs
• Objective: Next token prediction
• Configuration: DeepSpeed ZeRO-3, batch size 32,768 tokens, learning rate 1e-4

Stage 2: Supervised Fine-Tuning

• Dataset: ~700 instruction-response pairs
• Method: LoRA (rank 16, alpha 16)
• Training includes: translation, question-answering, summarization
• Loss computed only on response tokens

Stage 3: Preference Optimization

• Method: Iterative Reasoning Preference Optimization (RPO)
• Focus: Reducing glitching behavior and hallucinations
• Alpha parameter: 1.0

Evaluation

Translation Performance

The model was evaluated on a custom dataset with 100 sentences across 20 practical scenarios (healthcare, banking, education, agriculture, etc.) covering 31 Ugandan languages.

Average scores across 31 languages:

• chrF (xx→eng): 0.435
• chrF (eng→xx): 0.357
• BLEU (xx→eng): 20.625
• BLEU (en→xx): 7.598

Sunflower-32B achieves the highest accuracy in 25 of 31 languages when averaging bidirectional chrF scores.

Comparison with Other Models

Limitations

• Performance varies across languages based on training data availability
• Limited evaluation on tasks beyond translation and basic question-answering
• May generate content that reflects biases present in training data
• Not suitable for critical applications (medical diagnosis, legal advice) without human oversight
• Works best with text similar to training distribution

Bias and Ethical Considerations

The model was trained on data that includes historical texts, which may contain outdated views. Users should be aware that:

• Some translations use archaic language forms
• Training data filtering focused on removing harmful content but cannot guarantee absence of all biases
• Model outputs should be reviewed by speakers of the target language for critical applications

Citation

@misc{sunflower2025,
  title={Sunflower: A New Approach To Expanding Coverage of African Languages in Large Language Models},
  author={Akera, Benjamin and Nafula, Evelyn and Yiga, Gilbert and Natukunda, Phionah and Nsumba, Solomon and Muhanguzi, Joel and Namara, Janat and Sekalala, Imran and Walukagga, Patrick and Bainomugisha, Engineer and Mwebaze, Ernest and Quinn, John},
  year={2025},
  publisher={Sunbird AI}
}

Model Card Contact

For questions or issues, contact: [email protected]