YourSOTAGet a sample

HUMAN INPUT. REAL WORK. USEFUL DATA.

REAL PROJECT DATA FOR AI POST-TRAINING

Human instructions. The changes. The results.

What people asked for. How they changed the task. What the team made.

Human input → Multi-step instructions → Result & validation files

License real work from architecture, construction and smart-building projects to train and test your AI.

1+ TBof project data
Growingnew projects added
Request a sampleSee how it works

Contract-based sourcing. Licensing for your intended AI use.

Illustration connecting architectural drawings, a 3D building model and smart-building systems
DRAWINGS + MODELS + BUILDING SYSTEMS Illustration
A TASK, FROM START TO FINISHSelect a step ↓
Human inputWhat someone asked for.

The instruction that starts the work.

Human-written briefs and requirements explain the goal, the constraints and the expected result.

Briefs · requirements · constraints
Multi-step instructionsFollow-up requests. Changes to the first instruction.

See how the request changed.

Where history is retained, later messages and revised instructions show each change to the original task.

Messages · revised instructions · file versions
Result filesWhat the team made.

Open the work behind the instruction.

Models, drawings and documents show what was produced. Linking each result to its instruction is part of the agreed preparation.

BIM models · CAD drawings · project documents
Validation filesHow the result was checked.

See the checks and review notes.

Available checks, comments and acceptance records help assess the result. A finished file alone is not proof that it was validated.

Checks · review notes · acceptance records

History, file versions and review records are included where available. Each sample shows exactly what it contains.

ARCHITECTURECONSTRUCTIONSMART BUILDINGSReal work. Useful context.

01 / FOLLOW THE WHOLE TASK

The request is useful.
The story around it is richer.

A model needs more than a finished file. It needs to understand what the person wanted, what changed and how the work responded.

01

HUMAN INPUT / INSTRUCTION

“Here is what I need.”

The original brief, requirements and constraints. The human input that starts the work.

02

MULTI-STEP TASK INSTRUCTIONS

“Now change this part.”

Follow-up messages and revised requirements, where retained. See how instructions change over several steps.

03

RESULT & REVIEW

“Here is the work.”

Result files show what was made. Available validation files show what was checked or reviewed.

Multi-step task instructions with iteration over the initial instruction.
In plain English: a person asks for something, adds changes, and the work is updated.

A SIMPLE EXAMPLE

One request.
A change.
A file to inspect.

Illustrative workflow, not an actual conversation from the dataset.

  1. REQUEST

    “Create a room layout with space for a desk.”

  2. CHANGE

    “Keep the desk, but move it closer to the window.”

  3. RESULT

    A revised drawing or model, linked to the task.

  4. VALIDATION

    Available review notes explain what was checked.

02 / WHAT YOUR AI CAN LEARN

Follow the request.
Handle the changes.
Work with real files.

Select source files for your own pipeline, or agree on a package that links instructions to the work they produced.

INSTRUCTION FOLLOWING

Understand what people want.

Use human-written requirements to build tasks for domain-specific fine-tuning.

MULTI-STEP WORK

Respond when the task changes.

Where histories exist, test whether a model follows new instructions while keeping earlier context.

MULTIMODAL AI

Connect words to models and drawings.

Link text with BIM geometry, CAD drawings and project documents.

EVALUATION & BENCHMARKS

Check performance on real tasks.

Scope a pilot with reference files, review criteria and separate training and evaluation projects.

Task preparation, annotation and benchmark creation are separate scopes. Performance is measured in your pilot.

03 / REAL SOURCE MATERIAL

Made for real projects.
Ready for a closer look.

Technical briefs, BIM models, CAD drawings and project documents. More than 1 TB of industry data, with new projects added and targeted sourcing available to expand your coverage.

ONE EXAMPLE FILE55 MB

A single archive to illustrate the data.

THE WIDER DATA CORPUS1+ TB

More than a terabyte. New data added.

ONE EXAMPLE ARCHIVE

55 MBone sample archive

1 archive file · 10 source files inside · 4 formats

Includes human-written technical requirements, tracked edits to those requirements and BIM/CAD project files.

The documents in this example are in Russian. This example does not represent the full corpus or its language coverage.

IFCStructure & MEP models2 files · IFC2X3 exports
RVTNative BIM model1 file · Revit format
DWGArchitectural floor plans1 file · native CAD source
DOCXRequirements & project documents6 files · human-written project context

Tracked document edits are present in this example. Conversation trails, linked file versions and completed validation reports are checked separately for each proposed package.

04 / KNOW WHAT YOU GET

Start small.
See the data before you scale.

Request a representative sample and a data brief. Your research, engineering and legal teams can review the same proposed package.

01

What is included?

Project and file counts, measured volume, languages, formats and software compatibility.

02

What is connected?

Links between instructions and files. Available messages, revisions and review records. Any missing stages.

03

What has been checked?

File integrity, duplicates, missing references and known limits. How training and evaluation projects are separated.

04

How will it arrive?

Native files or separately prepared records. Discuss extraction, linking, translation, annotation and JSONL or Parquet exports.

05

How can you use it?

Source information, relevant permissions and license terms for your intended AI use.

06

What does the deal cover?

Sample and pilot scope, delivery, acceptance criteria, pricing and an optional supply schedule.

Bring your task and evaluation criteria.
We will discuss the sample that fits.

Download the buyer checklist

05 / WHERE THE DATA COMES FROM

Contract-based sourcing.
Clear terms for your use.

We obtain professional project data through direct agreements with data owners and agent-assisted acquisitions.

We discuss the source, relevant permissions and the license your team needs before production delivery.

Discuss licensing
01

A source your team can review.

Discuss the source party, contractual basis and relevant authorization, including the sourcing partner’s role where applicable.

02

Rights for the work you plan.

Define training, fine-tuning, evaluation or commercial model use. Agree on duration, territories, recipients and any redistribution or derived-data terms.

03

Privacy and confidentiality in scope.

Agree on review of personal information and third-party content, redaction, access, transfer and retention requirements.

04

Commercial terms in writing.

Discuss warranties, remedies, liability, exclusivity and treatment of trained models after license expiry.

06 / FROM A SAMPLE TO A SUPPLY

A simple way
to start working together.

A focused pilot, a corpus license or ongoing supply. Scope, volume, cadence and pricing are agreed individually.

01

Tell us the task.

Your model, target skills, languages and intended use.

02

Review a sample.

Inspect the files, task coverage and data brief.

03

Run a pilot.

Agree on preparation and measure fit in your pipeline.

04

License and grow.

Finalize terms, then scope more projects and optional updates.

07 / A FEW USEFUL ANSWERS

Questions before
the first conversation?

Is the data ready to train a model?

The starting point is professional source material. Extraction, linking, task assembly, annotation and normalized training records are separately scoped. We agree on what will be delivered and how it will be assessed.

Does every project include the full conversation?

No single package structure applies to every source. Conversation history, instruction revisions and file versions can be included where retained and permitted. The sample brief identifies what is present and what is missing.

Are result files the same as validation files?

Result files are what the team made. Validation files are checks, review comments or acceptance records showing how the result was assessed. A finished file on its own does not prove it was checked or accepted.

Which languages and formats are available?

The reviewed example contains Russian-language documents and IFC, RVT, DWG and DOCX files. Languages, formats and compatibility for your purchase are specified in its data brief. Translation and normalized exports can be discussed separately.

How much data can we buy?

Our industry corpus contains more than 1 TB of data and continues to grow. The 55 MB archive shown here is one example. We can source additional projects through contractual agreements and rights acquisition. Measured counts and volume are provided for the proposed purchase. Expansion targets and timing are agreed before commitment.

Can our legal team review the licensing?

Yes. Include your source and licensing requirements in the inquiry so we can discuss the acquisition route, supporting authorization information, restrictions and intended AI use before a production license.

How are price and exclusivity decided?

Terms depend on the chosen data, rights, preparation work and supply schedule. A pilot, a corpus license and ongoing supply can be discussed separately. Exclusivity, warranties and any indemnities require explicit agreement.

What does a linked task record look like?

We can scope records with an initial instruction, source references, changes in order, associated file versions and available review evidence. Missing stages and unknown review status should remain explicit. This is a target structure, not a claim that all source data is already converted.

08 / GUIDES FOR AI BUILDERS

Better data decisions.
Before the next training run.

Practical guides for founders and model teams: buy the right data, choose the right experiment and evaluate real task history.

Research-informed guides. Clear next steps.Explore all insights →

LET’S TALK ABOUT YOUR DATA

What should your AI
learn to do?

Tell us the task. We will discuss a sample, what it contains and the terms for using it.

Source & task coveragePreparation & deliveryProvenance & licensing

We confirm sample availability in discussion. An inquiry creates no purchase commitment.