What You'll Build

• A localized Qwen2.5 or Llama-3.2 model fine-tuned on company policies and FAQs
• A deterministic evaluation suite to ensure 100% adherence to company refund and password procedures
• An instant local REST endpoint served via Ollama and queried with Python

Step 1: Prepare Training Data

Format your customer support FAQ into structured multi-turn conversation JSONL:

cat > support_data.jsonl << 'EOF'
{"messages": [{"role": "system", "content": "You are a helpful support assistant for Acme Corp."}, {"role": "user", "content": "How do I reset my password?"}, {"role": "assistant", "content": "To reset your password, go to Settings > Security > Reset Password. You'll receive a confirmation email within 2 minutes."}]}
{"messages": [{"role": "system", "content": "You are a helpful support assistant for Acme Corp."}, {"role": "user", "content": "What is your refund policy?"}, {"role": "assistant", "content": "We offer a 30-day money-back guarantee on all purchases. To request a refund, go to Orders > Select Order > Request Refund."}]}
{"messages": [{"role": "system", "content": "You are a helpful support assistant for Acme Corp."}, {"role": "user", "content": "How do I contact emergency support?"}, {"role": "assistant", "content": "You can reach our 24/7 priority support team at priority@acme.com or by calling 1-800-ACME-HELP."}]}
EOF
  

Step 2: Initialize Project & Copy Data

moro init support-bot --template support-bot
cd support-bot
cp ../support_data.jsonl data/raw/
  

Step 3: Configure moro.yaml

name: "support-bot"
version: "1.0.0"

dataset:
  source: ./data/raw/support_data.jsonl
  format: jsonl
  mi_guard_threshold: 0.85

model:
  base: "Qwen/Qwen2.5-1.5B"
  quantization: "nf4"

training:
  learning_rate: 2.0e-4
  micro_batch_size: 2
  gradient_accumulation_steps: 8
  auto_heal: true
  

Step 4: Build & Train

# 1. Compile dataset with MI Guard
moro data build

# 2. Start self-healing training
moro train
  

Step 5: Run Evaluation Suite

Create an evaluation definition to verify accurate answers:

cat > eval/support_eval.yaml << 'EOF'
name: "support-eval"
tests:
  - id: "test_password_reset"
    prompt: "How do I reset my password?"
    expected_keywords: ["Settings", "Security", "Reset Password"]
    forbidden_terms: ["call the CEO", "I don't know"]
  - id: "test_refund"
    prompt: "Can I get my money back?"
    expected_keywords: ["30-day", "money-back"]
EOF

# Execute evaluation
moro eval run --checkpoint ./runs/best/ --suite eval/support_eval.yaml
  

Step 6: Deploy to Local Ollama Runtime

moro release export --run-id best --version v1.0.0 --quant Q4_K_M
moro release deploy --target ollama --version v1.0.0 --name support-bot
  

Step 7: Production Query Example (Python)

import httpx

def query_support_bot(user_question: str) -> str:
    """Send customer query to locally running Ollama instance."""
    response = httpx.post(
        "http://localhost:11434/api/generate",
        json={
            "model": "support-bot:v1.0.0",
            "prompt": user_question,
            "stream": False,
        },
        timeout=30.0,
    )
    return response.json()["response"]

# Example usage:
answer = query_support_bot("How do I reset my password?")
print("Support Bot Response:\n", answer)