TADHack 2025 synthetic customer-service vCons

43 synthetic customer-service calls in IETF vCon format (syntax 0.4.0, draft-ietf-vcon-vcon-core-02), built as demo data for TADHack 2025. The calls are set at Aquidneck Yacht Brokers, a fictional yacht brokerage, and are dated 2025-05-18 through 2025-05-24.

Synthetic data

Every conversation, name, phone number, email address, and order detail is generated by the vcon_faker pipeline. No real customer or agent appears. Each party is marked "validation": "synthetic", and every vCon carries a lawful_basis attachment whose proof mechanism records that it is synthetic demo data with no real data subject. The phone numbers and emails look real but were not checked against real subscribers, so do not dial or email them.

Overview

The dataset holds 43 calls between agents and customers, one vCon per call, covering marine-industry support scenarios. dataset.json is the manifest; the GitHub tag is the dataset version.

Conversation Types

1. Returns & Refunds

• Customers requesting returns for yacht equipment
• Processing refund requests
• Emotional customers (often expressing sadness about returns)

2. Shipping & Logistics

• Yacht transportation inquiries (e.g., Fort Lauderdale to Newport)
• Delivery status updates
• Shipping cost questions

3. Order Issues

• Wrong items received (e.g., yacht anchor instead of navigation system)
• Missing order investigations
• Order verification and corrections

4. Equipment Support

• GPS malfunction troubleshooting
• Navigation system issues
• Equipment compatibility questions

5. Business Services

• Yacht listing inquiries
• Brokerage service questions
• Pricing and commission discussions

6. Account Management

• Membership cancellations
• Billing inquiries
• Privacy and data concerns
• Contact information updates

7. Appointments & Scheduling

• Yacht viewing appointments
• Service scheduling
• Consultation bookings

Call Characteristics

• Duration: 43 to 77 seconds, about 58 seconds on average
• Call Disposition: All marked as "ANSWERED" with "VM Left" status
• Language: English
• Transcription Confidence: 99%
• Professional Tone: Agents maintain consistent, helpful demeanor
• Resolution Rate: Most issues resolved or appropriately escalated

Data Format

Each conversation includes:

• Audio recording (MP3, referenced by URL with a content_hash)
• Full transcript with speaker diarization
• AI-generated summary
• Participant metadata (names, roles, contact info)
• Call metadata (duration, timestamp, disposition)

Running the Viewer

The repository includes a Streamlit-based viewer for exploring the VCON conversations. To run the viewer:

1. Install the required dependencies:

pip install streamlit pandas

2. Run the viewer:

streamlit run vcon_viewer.py

The viewer will open in your default web browser and provides the following features:

• Browse conversations by day
• View conversation metadata and transcripts
• Listen to audio recordings
• Filter and search through conversations
• View detailed analysis and diarization

Typical Interaction Flow

1. Agent greeting with company name and agent introduction
2. Customer name verification
3. Issue description by customer
4. Information gathering (order numbers, email verification)
5. Resolution or escalation
6. Professional closing

Notable Patterns

• Customers frequently express emotions related to their issues
• Agents consistently follow verification protocols
• Marine industry-specific terminology used throughout
• Focus on high-value transactions typical of yacht brokerage

The dataset is meant for testing and demonstrating vCon tooling. The dialog is LLM-generated from a small set of scenarios, so it is not suitable for training models meant to reflect real customer behavior.

License

MIT. See LICENSE.