chewie-llama-3b

Chewie-Llama-3.2-3B

Chewie is a specialized medical Large Language Model (LLM) fine-tuned to assist Community Health Workers (CHWs) in sub-Saharan Africa. It is built on top of Llama 3.2 3B Instruct and optimized for:

1. Clinical Triage: Strict adherence to "Assessment -> Action -> Advice" protocols.
2. Danger Sign Detection: Identifying critical conditions (e.g., Pre-eclampsia, Sepsis) requiring immediate referral.
3. Bilingual Support: Fluent in both English and Swahili.

Model Details

• Architecture: Llama 3.2 3B (QLoRA Fine-tune).
• Dataset: electricsheepafrica/chewie-instruct (~3,100 clinical instructions).
• Training: 2 Epochs on A100 GPU using standard CHW guidelines (WHO/local MoH).

Usage

Installation

pip install torch transformers peft accelerate

Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "electricsheepafrica/chewie-llama-3b"

# Load Base Model + Chewie Adapter
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B-Instruct",
    device_map="auto"
)
model.load_adapter(model_id)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")

# Run Inference
prompt = """
<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a Community Health Worker (CHW). Provide safe, structured advice.<|eot_id|><|start_header_id|>user<|end_header_id|>

Mtoto ana homa kali na shingo ngumu. Nifanye nini?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Performance

• Protocol Adherence: 95.8%
• Referral Accuracy: 91.7%
• Benchmarks: Outperforms base Llama 3.2 on AfriMed-QA consumer queries (qualitative analysis).

Limitations

• Not a Doctor: This model is a decision support tool for trained CHWs, not a replacement for professional medical diagnosis.
• Scope: Focused on primary care and triage; may hallucinate on complex specialists topics (e.g., neurosurgery).

License

Apache 2.0