Conversational AI training data
Multi-turn dialogs written or role-played by native speakers, the way real customers talk: corrections, hesitations, changes of mind and local phrasing. Text or recorded speech, labeled with intents, slots and outcomes for training chatbots, voice assistants and support agents.
CustomerHola, quiero cambiar mi vuelo del viernes al domingo.
AgentClaro. ¿Me comparte su número de reservación?
CustomerEs ABC123… perdón, ABC132.
Delivered as JSONL, one dialog per line, with turn-level labels.
- Dialogs
- Customer support, sales, booking, banking, health and general chat
- Made by
- Native speakers writing or role-playing both sides
- Formats
- Text dialogs, or recorded speech with transcripts
- Labels
- Intents, slots, dialog acts, outcomes and sentiment
What conversational training data is
Conversational AI training data is a set of example dialogs between users and an assistant, often labeled with what the user wants (the intent) and the details they give (the slots). Chatbots trained on tidy, scripted conversations break when real people ramble, correct themselves or mix languages, so good data has to include those messy turns.
Translated dialogs rarely sound like real customers. We write and role-play conversations natively in each language, with scenarios and personas you agree.
What we create
From your scenarios, personas and label schema.
Written dialogs
Multi-turn conversations written by one writer from a scenario brief.
Role-played dialogs
Two people play customer and agent live, which gives more natural turns than one writer can.
Spoken dialogs
Role-plays recorded as audio, with transcripts and speaker turns. More on speech data.
Intent and slot labels
Turn-level labels to your schema, including dialog acts and the outcome of each conversation.
Utterance variations
Many ways to say the same request, written by different native speakers to avoid one person's habits.
Local and mixed language
Regional phrasing, slang and mixed-language turns, such as Spanish and English in US support chats.
How a dialog data project runs
You approve sample dialogs before volume work starts.
Scenarios
Use cases, personas, languages and the label schema.
Samples
A few dialogs per scenario and language for you to approve.
Production
Writers and role-play pairs create dialogs to agreed quotas.
Review
A second reviewer checks language, realism and labels.
Second review QADelivery
JSONL or your format, with scenario and language metadata.
Related AI data services
Questions about conversational data
What is chatbot training data?
Chatbot training data is a collection of example conversations and user requests, usually labeled with intents and slots, that teaches a chatbot to understand users and respond correctly.
Are the dialogs written or recorded?
Either. We write dialogs, have two people role-play them in text, or record spoken role-plays with transcripts, depending on what your model needs.
Can you use our intent and slot schema?
Yes. We label to your schema and guidelines. If you don't have one yet, we can draft it with you from your use cases.
Do you use real customer conversations?
We create new dialogs from scenarios, so no real customer data is needed. If you want us to label your own anonymized logs, we can do that under your NDA.
Request sample dialogs
Send your scenarios, languages and label schema, or book a call. We write sample dialogs for you to approve.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time