How to Sell Data to Model Labs

DataVendor is HUD's marketplace for selling code, data, HUD tasksets with their reinforcement learning (RL) environments to model labs. It gives software owners and experts a way to bring useful training material to the teams building AI models.

You can offer a codebase you already own or create exercises based on work you know well. On DataVendor, a ready package can become a listing. If your codebase is not ready to upload, you can register its description in Inventory and link the data later.

HUD helps you build and review environments where AI agents—systems that take actions in software—practice tasks. DataVendor is where you offer that work to buyers. This guide follows the path from choosing an asset to preparing a marketplace offer.

How to sell through the DataVendor marketplace

Start by deciding what you are ready to offer. If the files or tasks are prepared, a listing lets you present the package with pricing and delivery details.

If your codebase is not ready to upload, register it in Inventory. Describe the repository and add any metadata you already know. Link its data when you are ready, then create a listing. Registering metadata does not publish an offer or guarantee a sale.

Tier 2 vendors can also respond to buyer opportunities by submitting a new or existing listing that meets the request. The buyer purchases it through the normal checkout once it is live.

Choose the path before investing in a large project. A buyer who wants your raw code may have its own team to build tasks. Another may want you to deliver a working environment.

Choose what to offer

DataVendor supports several forms of supply:

• Code repositories: source files and the project history that explains them.
• Raw data ZIPs: datasets, documents, or other files packaged together.
• Tasksets: exercises with inputs, instructions, and a way to check results.
• RL environments: software where agents can attempt tasks and get feedback. They sell with the HUD tasksets that run on them, rather than as standalone listings.

The supply guide explains the options. Enterprise records can be uploaded as Raw data ZIPs.

For a software owner, start with a project that has a clear purpose and useful history. For an expert, start with a task you understand well enough to write instructions and judge the answer.

The buyer needs to see what the material could help a model do. For example, records from a reporting tool might support tasks about finding data, changing a query, or fixing an incorrect total. Explain which use your package supports today.

Prepare an example for buyer review

One complete example can show whether the data has enough context to use. It also helps you estimate the work needed to prepare the rest.

For a coding task, gather the bug report, the code before the fix, setup instructions, the accepted fix, and relevant tests. Explain how those records connect. A buyer can then see both the problem and how your team checked the result.

Keep the full reviewed archive for the buyer. If you turn it into an exercise, build the agent’s workspace from the starting version. Check that later commits, branches, tags, comments, and attached patches do not reveal the reference fix. Keep the final scoring checks outside the files the agent can change.

For a document task, gather the request, source files, expected result, and review method. Say whether the example is typical of the collection or one of a few complete cases. That helps a buyer judge how much work remains.

Check rights before sharing files

Review who owns the code or data and what the relevant agreements allow. Customer records, contractor work, and third-party software may each come with different limits.

Search the full package for passwords, personal details, and confidential material. For a repository, include old commits and linked tickets. For an environment, make sure it cannot reach live customer systems.

Removing names alone may not make records anonymous. The UK's ICO explains that people can remain identifiable through other details, even when direct identifiers have been replaced. ICO guidance

Have the planned use reviewed by a qualified adviser. Keep a delivery copy with a record of what was removed or changed, and make the limits clear to the buyer.

Use HUD when you need working tasks

If you are selling existing files, focus on explaining and checking them. If the offer includes tasks that agents can run, you also need a working setup and reliable scoring.

HUD provides tools for building and inspecting RL environments. You can use those tools to investigate failed runs, scoring mistakes, and shortcuts that earn credit without completing the task. See the HUD platform for the building workflow.

Start with basic checks. Run a known correct answer and a wrong one. Try an incomplete solution that should not pass. Confirm that each new attempt starts from the expected state and does not inherit changes from an earlier run.

Record what you tested and any remaining limits. This gives buyers evidence they can assess. It also helps you avoid promising that one task covers more than you have checked.

Create your marketplace offer

A DataVendor listing needs a clear description of the package, its price, and how the buyer will receive it. Publishing requires vendor setup and quality checks. The workspace Mutual NDA covers submission. Follow the listing guide when the package is ready.

Explain what is included in terms a new reader can follow. Name the skill or subject, the amount of material, the versions covered, and known gaps. Add setup notes for software and scoring notes for tasks.

Decide what support the price covers. Helping a buyer follow your setup instructions is different from rebuilding the environment for another software version. If you can provide extra work, describe it separately so both sides know what must be agreed.

Set a price and delivery plan

Price depends on what the buyer needs, how useful the package is, and the work left to do. Separate access to existing data from creating tasks, keeping software running, and sending updates.

Also define the license or sale terms. Agree on how the buyer may use the material, who may access it, and whether you can offer it to others. Make payment timing and the buyer's checks clear before starting custom work.

Catalog buyers can purchase a full listing at its listed price. A buyer can also purchase a listing submitted to its opportunity through the same checkout. Confirm seller payout timing separately from the buyer’s checkout payment.

DataVendor's buying guide explains the delivery paths. Repositories and bundles are downloaded; buyers copy HUD tasksets and their environments into their team. Eligible HUD taskset listings can offer a paid sample containing one task and its environment.

For your handoff, record the version delivered, setup steps, and checks performed. Say who will answer questions and how long support lasts. A published offer is one step toward revenue; a completed purchase is what makes it a sale.

Start with DataVendor

You can begin with an existing asset or an idea for new training tasks. The useful first step is a clear account of what a model lab could buy and what you would need to prepare.

Add assets to DataVendor Inventory to describe potential supply, or start a listing when your package is ready. If the work calls for a runnable environment, use HUD to build and review it before offering it to buyers.

Frequently Asked Questions

How do I sell data to model labs through DataVendor?

If the files or tasks are prepared, a listing lets you present the package with pricing and delivery details. If your codebase is not ready to upload, register its description in Inventory and link its data when ready. Tier 2 vendors can respond to buyer opportunities by submitting listings that meet the request.

What forms of supply does DataVendor support?

DataVendor supports code repositories, Raw data ZIPs, and HUD tasksets with their RL environments. Environments cannot be listed on their own.

When do I need HUD to sell data?

If you are selling existing files, focus on explaining and checking them. If the offer includes tasks that agents can run, you also need a working setup and reliable scoring. HUD provides tools for building and inspecting RL environments.

How do I keep a coding task from revealing the answer?

Build the agent's workspace from the starting version. Check that later commits, branches, tags, comments, and attached patches do not reveal the reference fix. Keep the final scoring checks outside the files the agent can change.

Does removing names make data anonymous?

Removing names alone may not make records anonymous. The UK's ICO explains that people can remain identifiable through other details, even when direct identifiers have been replaced.

When does a marketplace offer become a sale?

A published offer is one step toward revenue; a completed purchase is what makes it a sale. Catalog buyers can purchase a full listing at its listed price. A buyer can also purchase a listing submitted to its opportunity through the same checkout.