ALAMZ TECH SME Copilot (GGUF, Q4_K_M)
Built with Qwen. This model is a fine-tune of Qwen2.5-3B-Instruct and is therefore
governed by the Qwen RESEARCH LICENSE AGREEMENT (non-commercial use only; commercial
use requires a separate licence from Alibaba Cloud). Modifications made: QLoRA fine-tune
on an SME back-office / Nigeria-2025-tax corpus, then imatrix Q4_K_M quantization.
Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud.
All Rights Reserved.
An offline back-office copilot for African small businesses — invoices and quotes,
mobile-money (MoMo/M-Pesa) reconciliation, and Nigeria's 2025 Tax Reform Acts —
built for the 8 GB laptops with integrated graphics that SMEs actually own.
ALAMZ TECH's entry to the Africa Deep Tech Challenge 2026 (domain: corporate/enterprise).
- Base: Qwen2.5-3B-Instruct, QLoRA fine-tune (rank 32, all layers, all attn+MLP modules)
- Format: GGUF Q4_K_M with a domain-calibrated importance matrix — 1.93 GB
- Runtime: llama.cpp — CPU-only is fine
- Peak RAM: ~2.0 GB measured (audit-style container, 4 CPUs / 7.5 GB)
- Everything (data pipeline, training, quantization, evals) was built on one 8 GB M2 laptop
Why it exists
Nigeria rewrote its tax law in 2025 — after every mainstream base model's training data.
Stock models confidently quote repealed rates (5% VAT, ₦25M small-company threshold).
This model was fine-tuned on a hand-curated, grep-verified fact base built from the
OCR'd official Gazette (Nigeria Tax Act 2025 + Tax Administration Act 2025), with every
training number either confirmed verbatim in the Act or corroborated by 2+ professional
sources, and all arithmetic in the training data computed programmatically, never generated.
Measured on a 37-question adversarial fact eval (paraphrases, casual/Pidgin phrasings,
adversarial framings, greedy decoding):
24/37 (base-prior v1) → 34/37 (this model),
including: VAT 7.5% · small-company 0% CIT (≤₦100M turnover, ≤₦250M fixed assets) ·
standard 30% CIT · Development Levy 4% with the small-company exemption · the
professional-services exclusion. Full methodology and results in the
GitHub repo.
Run it
1# chat UI at http://localhost:8080
2llama-server -m alamz-tech-sme-copilot-Q4_K_M.gguf --port 8080 -c 2048
3
4# or one-shot
5llama-cli -m alamz-tech-sme-copilot-Q4_K_M.gguf -p "What is the current VAT rate in Nigeria?"
Try: "A customer paid NGN 127,500 by MoMo. They owe INV-114 (NGN 85,000) and INV-121
(NGN 42,500). Does this clear both?"
In the full product this model is paired with a deterministic finance module
(mobile-money statement parser, double-entry ledger, citeable tax-rule engine — same
verified fact base) that computes every figure; the model narrates. See the demo app in
the GitHub repo.
Limitations
- Nigeria-2025 depth is the specialty; other jurisdictions get general reasoning only.
- Like any small LLM it can err on multi-step arithmetic — the paired module exists
precisely for that; don't ship model-only math to production.
- Not professional tax advice. Confirm specifics with FIRS/NRS or a licensed accountant.
sha256 8ea4493dc50391a48cfc400a447c23dc50c584845c11ad0c999b7b030a5d773d — verify your download.
