stindardlogic/customer-service-sft-50k

Customer Service SFT (50K)

50,000 ShareGPT-format customer service conversations across 8 industries and 18 issue types. Each conversation includes a system prompt establishing the agent's role, authority limits, and policy constraints — training models to operate within defined boundaries while resolving issues empathetically and effectively.

Motivation

Customer service is one of the highest-volume LLM deployment contexts. Models need to balance:

• —Empathy with efficiency
• —Policy compliance with customer goodwill
• —Honesty when the answer is "no" with alternatives that still help
• —De-escalation of angry customers without being dismissive
• —Clear communication about what is and isn't possible

Most customer service datasets are either too scripted (robotic) or too vague (no real policy constraints). This dataset trains on realistic interactions where agents operate within real authority limits.

Dataset Description

50,000 conversations across 8 industries and 18 issue types:

Industry Distribution

Issue Type Coverage

Billing disputes, wrong/late orders, returns and exchanges, fraud and account security, technical support, service cancellations, complaint handling, policy explanations, price matching, escalation handling, de-escalation, feature requests, upgrade requests, and more.

Format

Standard ShareGPT with system prompt:

{
  "conversations": [
    {
      "from": "system",
      "value": "You are a customer service representative for ShopNow, an online retail platform. You are helpful, empathetic, and focused on resolving issues efficiently. You have authority to issue refunds, replacements, and courtesy credits for orders within the last 90 days."
    },
    {
      "from": "human",
      "value": "I ordered a blue medium t-shirt and you sent me a red large. This is ridiculous — I needed this for an event tomorrow and now I don't have anything to wear. What are you going to do about this?"
    },
    {
      "from": "gpt",
      "value": "I completely understand your frustration, and I'm truly sorry we sent you the wrong item — especially when you have a specific event tomorrow..."
    }
  ],
  "metadata": {
    "industry": "ecommerce",
    "issue_type": "wrong_item"
  },
  "id": "abc123"
}

Key Training Signals

1. Empathy before solution: Every response acknowledges the customer's situation before jumping to resolution steps. The ordering matters — customers who feel heard are more receptive to what comes next.

2. Specificity: Responses reference specific details from the customer's situation, not generic templates. "I can see the charges on March 3rd" vs. "I'll look into your account."

3. Authority transparency: Agents clearly state what they can and cannot do, and why. When a refund isn't available, alternatives are offered. "No" is always accompanied by what is possible.

4. Structured responses: Complex resolutions are formatted with clear steps, headers, and action items so customers know exactly what will happen.

5. De-escalation: Difficult customers (frustrated, angry, threatening to leave) are met with calm acknowledgment of the failure, not defensive responses.

6. Policy compliance: Agents operate within stated system prompt constraints. A final-sale item gets a policy explanation plus alternatives, not a policy violation.

7. Closing loop: Responses end with a next step, a confirmation request, or an explicit invitation to ask more — not a cold closure.

Difficult Scenarios Included

• —Customer demanding exception to explicit policy (non-refundable item)
• —Angry customer threatening cancellation after multiple failed contacts
• —Legitimate fraud claim requiring investigation
• —Medical emergency flight cancellation (exception process)
• —Billing error that requires cross-team coordination
• —Missing paycheck requiring urgent resolution

Use Cases

• —SFT fine-tuning for customer service chatbots
• —Training models to follow system prompt authority constraints
• —Tone and empathy training for enterprise chat deployments
• —Building product-specific support agents
• —De-escalation and difficult customer handling
• —Multi-industry support model training

License

Apache 2.0