Supervised Fine-Tuning on Ethiopian Commercial Datasets

Dataset curation, instruction tuning pipelines (LoRA/QLoRA), and synthetic data generation for Ethiopian commercial workflows.

1. Overview & Technical Architecture#

The sovereign model's domain expertise is achieved through targeted Low-Rank Adaptation (LoRA) fine-tuning on high-fidelity enterprise datasets.

Important

Offline Engineering Workflow: Supervised fine-tuning (SFT) is an asynchronous offline engineering process. SmartERP does not execute heavy backpropagation or GPU tensor training directly inside web workers or Odoo transactional runtimes. Training runs on external compute nodes, while SmartERP serves as the source of sanitized export datasets and the consumer of compiled model checkpoints.

1. Training Corpus Composition#

Statutory Legal Corpus (35%): Full-text Ethiopian Federal Negarit Gazeta proclamations, directives, and court precedents.
Accounting & ERP Taxonomy (40%): Standard charts of accounts, voucher workflows (Model 19, 20, 21), and tax calculation cases.
Bilingual Parallel Text (25%): English-Amharic aligned commercial contracts, tender specifications, and business correspondence.

2. LoRA Hyperparameter Configuration#

{
  "lora_r": 64,
  "lora_alpha": 128,
  "lora_dropout": 0.05,
  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
  "learning_rate": 2e-4,
  "lr_scheduler_type": "cosine",
  "warmup_ratio": 0.03,
  "per_device_train_batch_size": 4,
  "gradient_accumulation_steps": 8
}

2. Prerequisites & Infrastructure Requirements#

Runtime Environment: Python 3.12+, PostgreSQL 16 with pgvector extension.
Security Access Group: Ashewa AI / Administrator (Technical ID: muk_ai.group_ai_manager)
Hardware Profile: Minimum 16 GB System RAM (CPU inference) or NVIDIA GPU with 8+ GB VRAM.

3. Step-by-Step Configuration & Verification#

Step 1: Navigating to Model Management#

Access your SmartERP instance at https://smarterp.et/odoo.
Open Settings → Technical → Ashewa AI → Model Registry.
Select the custom model profile.

{screenshot:50-fine-tuning-on-ethiopian-business-data-step1-navigation}

Step 2: Testing Endpoint Health & Response Verification#

Click the Run Diagnostic Health Check button.
Verify token generation speed (tokens/sec) and latency benchmarks.
Review sample response in the diagnostic console.

{screenshot:50-fine-tuning-on-ethiopian-business-data-step2-verification}

4. Advanced Configuration Points#

Automated Tasks: muk_ai.ir_cron_model_health_monitor runs every 15 minutes.
Multi-Tenant Boundaries: Custom model fine-tuning adapters are strictly segregated per tenant database.

5. Comprehensive Troubleshooting Matrix#

Symptom / Error	Probable Root Cause	Resolution Action
`Connection refused: Model server unreachable`	Inference daemon stopped or port blocked by firewall.	Restart model service (`systemctl restart vllm`) and verify firewall rules.
`Token limit exceeded on input prompt`	Input context exceeded window size.	Enable context truncation in model profile settings.

Last updated: 2026-10-30