Tasks with clear rules
A request, its context, permitted actions, and success criteria. Include mixed-script identifiers, bilingual documents, and corrections that change the action.
Bayanat Agents
Arabic training data for agents that use tools.
Bayanat Labs (Riyadh) supplies Arabic agent training data: license trajectories with tool calls and outcomes, or commission executable tasks around your APIs from a team of 50,000+ expert annotators.
اقرأ بالعربيةWhich Arabic tool-calling or agent data fits your Saudi or UAE project?
Start from what your agent has to do. Ready trajectories can be licensed alongside Arabic Smartphone Speech; executable environments are built to order around your tools, so they are separate offers. The first two rows are specific to Saudi Arabia and the UAE; the rest apply in any market.
| Your project | Start with | Check before you commit |
|---|---|---|
| A booking or service agent for Saudi users that turns Saudi Arabic dates and times into actions | Custom executable environment Eight trajectory checks before you accept data | A reference time and the Asia/Riyadh time zone (UTC+3) in every starting state, each record’s Arabic variety labeled Saudi rather than Gulf, a confirmation before each state-changing call, and a check that reads the final booking, order or record. |
| An agent for UAE users, or one that serves both countries | Custom executable environment What Saudi and UAE teams add to the spec (in Arabic) | Asia/Dubai (UTC+4) recorded in every record, examples where a user names a city in the other country, Emirati Arabic named in the scope, an agreed format for amounts in AED or SAR, and where the data will be hosted: in-region is the default, with private-cloud and on-prem options. |
| Fine-tuning now on ready-made Arabic agent trajectories | Trajectories that can be licensed alongside Arabic Smartphone Speech | Versions, counts and tool schemas, which are on request; whether the trajectories can be licensed without the audio; and a sample trajectory, where one is available (request, tool calls, observations and outcome), before you choose a set. |
| Single-step function calling only: choosing a function and filling its arguments | Public Arabic function calling datasets compared | The license on each dataset card, whether a listing copies a set you already count, and exact argument match rather than function choice alone: in the AISA-ArabicFC pilot baselines, function accuracy reaches 0.982 but argument exact match tops out at 0.541. |
| Teaching an agent to recover from a failed call or a user correction | Custom failure and recovery tasks | Failure examples with a failure category attached, corrections that change the action, and negative controls: a run whose last message claims success while the state did not change must be scored as a failure. |
| Checking a delivered trajectory set before you accept it | Runnable state check (Python, CC BY 4.0) | That the check replays the state-changing calls, reads the final state and checks the order of calls against the confirmation rule, so it passes a correct run and fails both negative controls. |
| Evaluating an agent you have already built | Bayanat Bench Arabic Agent Reliability Lab (methods beta) | Whether the task checks what actually changed, not only what the reply says. The lab describes a method from a synthetic commerce sandbox; it is not a certification or a score for your agent. |
Arabic agent-trajectory sets are also available to license with our smartphone speech dataset. Ask us about the available versions and counts.
Is Bayanat Agents the right fit?
A good fit if
- You are training agents that use tools in Arabic and need trajectories with tool calls, observations and outcomes.
- You need an executable environment that checks the final state, not just the agent’s reply.
- You want failure and recovery examples, including corrections that change the action.
Choose something else if
- You only need to evaluate an agent you already built. Start with Bench.
- You need a production-safety certification. Our research is a methods beta, not a certification.
What you can check before you license
We separate what is confirmed on this page, what we send on request, and what is agreed for each project.
| Item | Status | Details |
|---|---|---|
| Licensable trajectories | Confirmed | Existing Arabic agent trajectories can be licensed alongside our Arabic smartphone speech dataset. Versions, counts and tool schemas on request. |
| Executable environments | Per project | Custom mock APIs, documents, databases or sandbox applications with agreed starting states. |
| Checks of the result | Per project | Each custom task verifies the final booking, order or record, with difficulty, language variety and failure category attached. |
| Sample trajectory | On request | Where one is available: request, tool calls, observations and outcome, before you choose a set. |
| Research method | Confirmed | Our Arabic Agent Reliability Lab is a methods beta that checks final state with evidence receipts in an open synthetic sandbox. |
What we can build together
Executable environments
Custom mock APIs, documents, databases, or sandbox applications give agents somewhere to act and make results reproducible.
Demonstrations and recovery
Tool calls and observations show successful paths. Failure examples and corrections teach recovery when a first attempt goes wrong.
Checks of the result
Verify the final booking, order, or record. Attach difficulty, language variety, failure category, and environment version to the task.
How the license covers each use
Training rights do not automatically include display, retrieval or redistribution. Each use is set in the agreement.
Model training
Permitted use is set in the agreement for the trajectory set you license.
Custom environments
Built on mock APIs, documents, databases or sandboxes defined with you, with the permissions and confirmations you specify.
How we put the data together
Bring a real task
Start with an existing trajectory set or describe an operation your agent needs to perform.
Define the boundaries
Identify permissions, required confirmations, and handoffs before we create demonstrations.
Inspect a complete run
For custom environments, review setup, actions, observations, and result checks together before expanding the task library.
For example: Rescheduling an appointment
An Arabic request gives an ambiguous date, then changes the time. The agent clarifies, keeps the requested time zone, checks availability, and confirms the completed booking.
Scope your request
Send us these details and we’ll send the available specifications, the licence wording, and a sample where one is available, so your team can check the fit before licensing. If you need annotation added, we scope a pilot with gold-standard QA and a quality report, usually within two weeks.
- The operation your agent must complete
- Tools or APIs involved
- Required confirmations and handoffs
- Languages and dialects
- Timeline
Common questions
What is an agent trajectory?
It records the steps an agent takes: the request, tool calls, observations, and outcome. It gives a training system examples of how to act, including how to recover from errors, rather than only how to describe an action.
Does every trajectory set include an executable environment?
No. Existing trajectory data and custom executable environments are separate offerings. If your team needs repeatable runs, we can build an environment with agreed APIs, starting states, and checks for success.