Skip to content

Train and test your model
on real work.

Tell us the task your model needs to perform. We find datasets with the right records, check coverage with their owners and help you test available samples before you license the data.

Start with one research task and what a sample would have to show.

Grainy editorial image of work moving between a notebook and keyboard

Business data.
And beyond.

Real-world collections for teams building and evaluating AI. Explore where your data fits.

The records behind how work gets done.

A whole-company archive or selected systems. Connected records can follow a customer request through decisions, work and its outcome.

  • Email, chat & customer support
  • Invoices, transactions & CRM
  • Project files, tickets & knowledge bases
  • Work orders, service & delivery history

Start with your sector, team size, systems and years of history. The records can still be in the tools you use today.

Microsoft 365Google WorkspaceCRM & work tools

Check whether a sample supports your task.

In this example, the job ID connects an estimate, a drawing and an invoice. A researcher can test whether a model follows the right revision through all three records.

Structured requirements

Task, category, required systems, geography, formats, volume, coverage and acceptance criteria.

Which criteria are essential and which are preferences.

Dataset specification

Contents by system, available amounts and dates, record relationships, provenance and known gaps.

Whether the material is relevant enough to inspect.

Sample review

Available sample files, supporting context and requests for missing records or complementary information.

Whether the data passes your own technical review.

Licensing & delivery scope

Permitted use, authority, exclusivity where available, price, delivery formats and conditions to resolve with the owner.

What can be licensed and on which terms.

Coverage varies by dataset. Unknown, estimated and supplier-reported information must remain distinguishable from verified facts. Suitability for an RL environment or evaluation task requires your technical review.

One job, across three systems

Illustrative records · invented IDs
SourceRecordWhat connects it
EstimatorEstimate EST-42Job J-042 · accepted revision 2
Project folderDrawing D-42Job J-042 · revision 2 built
AccountingInvoice INV-42Job J-042 · references EST-42

A reviewer can follow the accepted design into the drawing and invoice. A shared job ID makes the join testable.

Source for the capability you want to build.

Learn from real work

You need operational records that preserve context across conversations, tasks, documents, code and outcomes.

Go deeper in a domain

Your research calls for particular formats, modalities or metadata in healthcare, CAD, robotics, games or another specialist field.

Evaluate a specific task

You need to inspect the source material, linkage and completeness before deciding whether it can support your training or evaluation work.

Editorial illustration of an engineer working between a notebook and a computer

Trace a decision from the first record to the outcome.

If your model has to reconcile an invoice, it needs the accepted estimate and the job it belongs to. We check for those links and ask the owner for missing records.

Connected business records

Communications, documents and operational history across company systems, with the links between them where available.

Domain-specific depth

Healthcare, engineering and CAD, robotics, games and more. Search for the actual contents, formats and history your work needs.

Samples before a decision

Review available samples, inspect the structure and identify gaps. Request the missing records or details in your company workspace.

A fit for your use case

Define the source, rights and delivery requirements. Agree the scope of a licence with the data owner before proceeding.

Start with what you want AI to do.

Give us a specific brief. Your team can browse and save datasets in a shared workspace, or use the API and MCP from your own agent.

Prepare your brief
Capability
The task, domain or behaviour you want to train or evaluate.
Contents
Required systems, records, formats, volumes and history.
Quality
Linkage, annotations, metadata and sample acceptance criteria.
Licence
Intended use, access requirements, budget and timeline.

From a research brief to the right data.

  1. Define the requirements

    Use a structured buyer request to separate essential criteria from preferences.

  2. Review the evidence

    Compare matching inventory, download available samples and collect any missing information.

  3. Agree the licence

    Work through scope, rights, price and delivery with the data owner. The terms follow the actual dataset.

Price the actual scope and rights.

Dataset coverage, intended use, exclusivity, preparation and delivery affect commercial terms. We work through those details with the owner. A catalogue entry or sample does not by itself grant training rights or promise a model-performance improvement.

Your questions,
answered.

Can we inspect samples before licensing?

Available samples can be downloaded through the buyer workspace when access conditions permit. If a sample is missing, your team can request it. Sample availability and any restrictions depend on the dataset.

Can we specify requirements beyond the category?

Yes. Define the systems, formats, coverage, volumes, linkage, metadata and acceptance criteria that matter. Separate hard requirements from preferences so a relevant category is not mistaken for a complete match.

What if information is missing?

Record the specific gap against the dataset. Maryah can collect the additional detail or files from the owner and add them to the same record, rather than create a separate version for each buyer.

Are these ready-made RL environments?

Not necessarily. Some datasets can provide material for tasks and environments, but raw records are not the same as an executable environment with tools and a verifier. Review what each package actually includes.

Can our agent work with the catalogue?

The buyer workspace supports API and MCP access using the company's access permissions. Your agent can review available information and work with structured requests; the workspace remains available for manual review.

Is the catalogue public, and are all rights the same?

Catalogue access is controlled through a Maryah password or an authorized buyer workspace. Rights are agreed for the particular dataset and use case; access to a listing does not create a uniform licence across the catalogue.

What does your model need to learn?

Send a task, the records it requires and what a sample must show. We’ll review available datasets against those criteria and identify what we need to source from owners.

An existing RFP works too.