GeoNjunge/adtc-2026-fraud-auditors

Llama-3.2-3B-ADTC-Swahili-Auditor-Q4_K_M (GGUF)

Model Summary

Llama-3.2-3B-ADTC-Swahili-Auditor-Q4_K_M is a quantized, domain-adapted instruction model fine-tuned for bilingual health insurance fraud detection and clinical claim auditing across East Africa. It is designed for low-resource, CPU-only edge deployments, operating within strict memory constraints (~3.45 GB peak RSS).

The model performs joint analysis across:

• Bilingual Clinical Narratives: Verifying diagnosis consistency across English and Swahili clinical notes.
• Document Metadata: Detecting authoring tool anomalies (e.g., claims created via graphic suites rather than hospital EMR systems) and timestamp mismatches.
• Billing Validation: Cross-referencing claim billing items against standard regional tariff expectations (KES) and demographic alignment.

---

Model Architecture & Quantization

• Base Model: Llama 3.2 3B Instruct
• Parameters: 3.21 Billion
• Format: GGUF (4-bit Medium Quantization — Q4_K_M)
• Context Length: Up to 131,072 tokens
• Primary Languages: English (en), Swahili (sw)

---

Technical Specifications & Telemetry

Evaluated on standard budget hardware (Intel Xeon CPU @ 2.20GHz, 12.7 GB RAM, No GPU) using adtc-profiler:

Metric	Value
Quantization Type	GGUF Q4_K_M
Peak RAM Footprint (RSS)	3,449.77 MB (~3.45 GB)
Steady-State RAM (RSS)	3,339.38 MB
Generation Speed	3.16 tokens/sec
First Token Latency	65,238 ms (512 prompt tokens)
Benchmark Accuracy (arc_easy)	70.0% (acc_norm)

---

Usage

Local Execution with llama.cpp

You can run this model directly on CPU using standard llama.cpp builds:

./llama-cli \
  -m ./model/llama-3.2-3b-instruct.Q4_K_M.gguf \
  -p "### CLAIM SUBMISSION REPORT\n[METADATA]\n- CREATION_DATE: 2026-05-01\n- MODIFICATION_DATE: 2026-04-20\n- AUTHOR_SOFTWARE: GRAPHICS_DESIGN_SUITE\n[CLAIM_DATA]\n- PATIENT_GENDER: MALE\n- CLAIMED_DIAGNOSIS: Prostate Surgery\n- REQUESTED_PAYMENT: KES 950,000/=\n[CLINIC_NOTES]\nMgonjwa alikuja kwa ushauri wa afya ya ngozi. Hakuna matibabu ya upasuaji yaliyofanyika.\n\n[ANALYSIS_INSTRUCTION]\nTathmini data ya madai na maelezo ya matibabu hapo juu. Bainisha makosa (anomalies), toa maelezo, na uweke kiwango cha hatari (risk level)." \
  -n 256 \
  -c 2048 \
  --temp 0.2