Mwamuzi 0.8B (Decision & Guardrail Routing Model)

Mwamuzi ("The Arbiter / The Decider" in Kiswahili) is a high-speed, 0.8-billion parameter multi-task decision and guardrail routing model specifically fine-tuned for East African civic, public administration, financial, and digital public infrastructure applications.

Based on vllm-sr/Vela-2.0-0.8B, Mwamuzi resolves the zero-shot English prior bias on Swahili queries while maintaining native sub-200ms latency.

---

🎯 Benchmark Performance on Unseen Data (Held-out Test)

Evaluated against unseen real-world East African citizen queries:

Decision Dimension	Zero-Shot Baseline	Mwamuzi (Fine-Tuned)	Lift / Improvement
Swahili Language Identification (language)	40.98%	81.97%	+40.99% (Doubled)
Domain Routing (domain)	55.74%	61.48%	+5.74%
Query Complexity (complexity)	41.80%	54.92%	+13.12%
Personal Data / PII Guard (contains_pii)	96.72%	97.54%	+0.82%
Harmful Request Filter (harmful_request)	94.26%	95.08%	+0.82%
Crisis & Emergency Alert (crisis)	87.70%	92.62%	+4.92%
Jailbreak Defense (jailbreak)	91.80%	90.98%	Stable
Compound Exact Match (All 7 Decisions Correct)	16.39%	24.59%	+50% Relative
Inference Latency (p50)	168 ms	~170 ms	Sub-200ms native

---

🚀 Quickstart Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download

# 1. Load Base Vela Model
base_model_id = "vllm-sr/Vela-2.0-0.8B"
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

# 2. Download and Load Mwamuzi Head Weights
weights_path = hf_hub_download(repo_id="pawa-ai/Mwamuzi", filename="vela_head_finetuned.safetensors")
head_weights = load_file(weights_path)
model.head.load_state_dict(head_weights)
model.eval()

print("Mwamuzi is ready for sub-200ms East African decision routing!")

---

🏛️ Supported Civic Domains

• Government Services: NIDA IDs, Passports, Land Registry (Ardhi), BRELA Business Registration.
• Finance & Tax: Tanzania Revenue Authority (TRA TIN/Tax), Public Financial Systems (GePG).
• Mobile Money & Banking: M-Pesa, Airtel Money, Tigo Pesa, HaloPesa, Bank Accounts.
• Healthcare & Emergency: NHIF Insurance, Clinics, Emergency reporting.
• Agriculture & Livestock: Pest control, Crop markets, Farming weather.

---

📦 Checkpoint Details

• Architecture: Qwen3.5 Text Backbone + Multi-Task Bilinear CandidateHead
• Checkpoint File: vela_head_finetuned.safetensors (~4.2 MB)
• Trainable Parameters: 1,053,185
• License: Apache 2.0