Agent data and the callers to test it on.
Conversations between people and agents, recorded or simulated, with every tool call in the trace. Contributors play the caller with a persona and a task; simulated callers built on Svara add accents, interruptions and a noisy line, so the agent is tested on what it will meet.
When you need it
- You are building a voice agent for Hindi or Tamil callers and need conversations where the caller interrupts, changes their mind and mixes languages.
- You need tool-use trajectories with the right calls and arguments, judged by people, not only by whether the code ran.
- You want a regression suite of callers you can run against every release.
What one record looks like
An example task, as the rater or reviewer sees it. Try it.
- Mera balance kya hai?
- get_balance(account)120 ms
- आपका बैलेंस बारह हज़ार चार सौ पचास रुपये है।340 ms
Run replays the turns in order with the latency of each step.
- turns
- list
- User, agent and tool turns in order, with timestamps and audio for spoken turns.
- tool_calls
- list
- Name, arguments and result for each call.
- task and outcome
- text and label
- What the caller wanted and whether the agent achieved it, judged by a reviewer.
- persona
- object
- Language, accent, patience, background noise, for simulated callers.
- issues
- labels
- Invented fields, wrong tool, talked over the caller, and the rest of the checklist.
How Lokah builds it and checks it.
- 01
Tasks and personas
Written with you from real intents; personas cover the languages, accents and patience levels your callers have.
- 02
Real callers
Contributors call the agent with a task card, on their own phone. Every conversation is transcribed and the outcome judged.
- 03
Simulated callers
Svara voices with a persona, an accent, interruptions and line noise, driven by a script or a caller model, for volume and repeatability.
- 04
Delivery
JSONL trajectories with audio, the tool-call schema, and the caller suite as a runnable set.
How it arrives
- JSONL trajectories
- Turns, tool calls, outcome and issues per conversation.
- Audio
- 16 kHz per spoken turn, or the whole call on two channels.
- Tool-call schema
- The functions and argument types the traces use.
Other layouts on request. Catalogue layouts are described on formats.
Pairs with it
From Lokah, and why.
- Full-duplex call-centre conversations
How real agents and customers take turns, the target for a voice agent.
- Hindi call-centre conversations, insurance
Real intents and phrasing to write tasks from.
- Tamil two-channel customer-service calls
Real Tamil customer calls, with the phrasing and turn-taking a Tamil agent must handle.
Questions, answered.
Also from Lokah
Start a collection.
Tell us what you want to collect and from whom. We reply with how we would run it.
- Speech, images, documents, feedback or annotation
- A new collection, or a dataset from the catalogue
- Licensing and data handling
Prefer to reach out directly? Write to hello@kenpathlabs.com.