electricsheepafrica/chewie-1.2

Chewie 1.2: African Community Health Worker AI Assistant

Chewie 1.2 is a medical AI assistant fine-tuned specifically for Community Health Workers (CHWs) in Africa. Built on Google's MedGemma-4B, it provides structured clinical guidance following a consistent Assessment → Action → Advice format.

Model Details

Model Description

Chewie 1.2 is a LoRA adapter fine-tuned on top of google/medgemma-4b-it using a curated dataset of 11,880 multilingual medical conversations designed for community health settings in Africa.

• Developed by: Electric Sheep Africa
• Model type: Causal Language Model (LoRA Adapter)
• Language(s): English, Swahili, Hausa, Yoruba, Amharic, Zulu, French
• License: Apache 2.0
• Finetuned from: google/medgemma-4b-it

Model Sources

• Repository: electricsheepafrica/chewie-1.2
• Previous Version: electricsheepafrica/chewie-llama-3b

Intended Use

Primary Use Cases

• Clinical Decision Support: Helping CHWs assess symptoms and determine appropriate actions
• Triage Assistance: Identifying danger signs that require immediate referral
• Health Education: Providing patient-friendly explanations and preventive advice
• Multilingual Support: Serving diverse African language communities

Target Users

• Community Health Workers (CHWs)
• Primary healthcare providers in resource-limited settings
• Health education programs
• Mobile health (mHealth) applications

Out-of-Scope Use

• NOT for direct patient diagnosis - Always requires human clinical oversight
• NOT a replacement for professional medical care
• NOT validated for emergency/critical care decisions
• Should not be used without proper clinical supervision

How to Get Started

Installation

pip install transformers peft bitsandbytes accelerate

Inference Code

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

# Load base model with 4-bit quantization
base_model_id = "google/medgemma-4b-it"
adapter_id = "electricsheepafrica/chewie-1.2"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    attn_implementation="eager",
    device_map="auto",
)

# Load adapter
model = PeftModel.from_pretrained(base_model, adapter_id)
tokenizer = AutoTokenizer.from_pretrained(adapter_id)

# Generate response
def chat(prompt):
    messages = [{"role": "user", "content": prompt}]
    inputs = tokenizer.apply_chat_template(
        messages, 
        add_generation_prompt=True, 
        return_tensors="pt"
    ).to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            inputs,
            max_new_tokens=512,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
        )
    
    return tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)

# Example usage
response = chat("What are the danger signs in a child with diarrhea?")
print(response)

Training Details

Training Data

• Dataset Size: 11,880 examples
• Languages: English, Swahili, Hausa, Yoruba, Amharic, Zulu, French
• Domain: Community health, primary care, maternal/child health, infectious diseases
• Format: Structured Assessment → Action → Advice responses

Training Procedure

• Base Model: google/medgemma-4b-it
• Method: LoRA (Low-Rank Adaptation)
• Framework: Hugging Face TRL (SFTTrainer)

Training Hyperparameters

Parameter	Value
LoRA Rank (r)	16
LoRA Alpha	16
LoRA Dropout	0.05
Target Modules	all-linear
Learning Rate	2e-4
Batch Size	1-2
Gradient Accumulation	8-16
Epochs	1
Precision	bf16/fp16
Optimizer	AdamW (fused)

Training Results

Metric	Value
Final Training Loss	1.08
Final Validation Loss	1.13
Mean Token Accuracy	72.5%
Training Time	~1 hour

Compute Infrastructure

• Hardware: NVIDIA A100 40GB / T4 16GB (compatible with both)
• Cloud Provider: Google Colab
• Training Framework: Transformers, PEFT, TRL, BitsAndBytes

Evaluation

Sample Outputs

Input: "What are the symptoms of malaria?"

Output:

**Assessment:** Malaria can cause various symptoms, including fever, chills, 
sweating, headache, muscle aches, nausea, vomiting, and fatigue. In severe 
cases, it may lead to anemia, kidney failure, or brain damage.

**Action:** If you or someone you know has a high fever (above 102°F or 39°C), 
persistent headaches, or unusual weakness, seek medical help immediately at 
a clinic. These could be danger signs for severe malaria.

**Advice:** To prevent malaria, use insect repellent when outdoors, sleep 
under mosquito nets, and take preventative medications if recommended by 
your healthcare provider.

Limitations and Risks

Known Limitations

• Model outputs should always be verified by qualified healthcare professionals
• May not reflect the most current medical guidelines
• Performance may vary across different African languages
• Not trained on region-specific drug formularies or protocols

Ethical Considerations

• This model is intended as a decision-support tool, not a replacement for clinical judgment
• Healthcare providers should use their professional expertise alongside model outputs
• Patient safety must always take precedence over model recommendations

Citation

@misc{chewie-1.2,
  author = {Electric Sheep Africa},
  title = {Chewie 1.2: African Community Health Worker AI Assistant},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/electricsheepafrica/chewie-1.2}
}

Model Card Contact

• Organization: Electric Sheep Africa
• Issues: Please open an issue on the model repository

---

Framework Versions

• PEFT: 0.18.0
• Transformers: 4.x
• TRL: 0.x
• BitsAndBytes: 0.x
• PyTorch: 2.x