Skip to content

For Acquisition Partners

Business datasets for training and evaluation.

SourceX helps you source relevant company records, agree licensing terms and coordinate preparation and delivery through one relationship.

01 / Example data request

Data requestIllustrative

Multi-turn insurance claim workflows

TextStructured dataUS2018–presentDe-identified
Submit a data request

Candidate suppliers

  • Regional P&C carrierQualified
    92% fit
  • Claims TPASample under review
    78% fit
  • Independent adjusting firmRights review
    61% fit

02 / How sourcing works

We manage sourcing and licensing for you.

  1. Find

    We look for businesses in the supported countries that hold the data you described. Nothing is contracted until a supplier agrees.

  2. Assess

    We check the data and the supplier's licensing permissions.

  3. Agree

    Pricing and allowed uses are agreed in a license.

  4. Transact

    We coordinate delivery and payment.

  5. Manage

    We handle future purchases and the supplier relationship.

Supported countries: U.S., Canada, EU, UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, UAE, Saudi Arabia.

See what leading AI developers build, browse Acquisition Partners by industry or search licensing terms.

Search humanoid training data

Find guides on task demonstrations, procedures, objects, inspection, maintenance, warehouse, manufacturing and procurement. Browse all

03 / Example demand

The kinds of data we source

Illustrative examples of the kinds of datasets businesses in the U.S., Canada, the EU, the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia hold. These are categories we source, not inventory already under contract.

LegalExample demand

Legal reasoning workflows

Seeking
Complex legal workflows and attorney work product
Format
Documents + structured metadata
Geography
US
FinanceExample demand

Accounting workflows

Seeking
Month-end close, reconciliations, and review notes
Format
Spreadsheets + documents
Geography
US / UK
ConstructionExample demand

Construction project records

Seeking
Plans, RFIs, change orders, and inspection outcomes
Format
Documents + images
Geography
North America
InsuranceExample demand

Insurance claims workflows

Seeking
Claim files with adjuster decisions and outcomes
Format
Documents + structured data
Geography
US
SupportExample demand

Customer service conversations

Seeking
Multi-turn support threads with resolutions
Format
Text + audio
Geography
Global, multilingual
EngineeringExample demand

Software engineering histories

Seeking
Issues, pull requests, and review discussions
Format
Code + text
Geography
Global

04 / Rights and use

Clear rights. Defined use.

We coordinate dataset-specific licensing terms, so each dataset arrives with documented rights and a license that defines what you can do with it.

You control the scope

Sourcing starts from your requirements: domain, volume, language, format and timing. Nothing is licensed outside the scope you agree.

Preparation to specification

Each dataset follows a dataset-specific preparation specification, agreed before delivery.

Provenance and rights documentation

Where the records came from and the supplier's licensing permissions are documented before a license is signed.

Permitted uses

We coordinate dataset-specific licensing terms that set allowed uses, duration and restrictions.

Delivery standards

Format, quality checks and delivery method are agreed for each dataset.

Privacy processing

Personal details are removed before delivery, and the privacy processing applied is documented for each dataset.

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Acquisition Partner guides to sourcing and licensing AI data →

How do you turn a model failure into a data request SourceX can source?

Describe the failure as a business task, then name the context, actions and outcomes that records of that task must contain, and give each one its own field in the request form. A symptom such as the agent mishandles refunds gives the Find step nothing to look for; a task with named systems and outcome labels does.

The request form already has a place for each part. Choosing Workflow datasets opens a workflow specification: business task, workflow type, input or context required, desired action trajectory, required outcome labels and evaluation criteria. Volume, format, geography, language, historical timeframe, licensing and de-identification follow for every request. Datasets are sourced on request rather than held in stock, and sourcing starts from what you write in those fields.

Deciding which failure deserves data, and whether more examples can fix it at all, comes first, and our guide to post-training task selection covers that protocol. The worksheet below picks up where it ends, using an illustrative support agent that grants refunds outside policy.

How do you turn a model failure into a data request SourceX can source?
Part of the failure write-upRequest form fieldIllustrative entryWhy a supplier can act on it
The task the model gets wrongBusiness task; workflow typeDecide refund requests against the written refund policy; case resolutionIt names work a company does daily and records in its tools
What the model saw at the timeInput / context requiredTicket thread, order history and the policy version in force on the ticket dateIt lists the records every case must carry
What a correct handler didDesired action trajectoryCheck order status, apply the policy rule, then approve, refund in part or escalateThe steps appear in a help desk's case history
How each case endedRequired outcome labelsRefund issued, refused or escalated, and whether the customer reopened the caseOutcomes the source system already stores need no new labeling
How you will judge fitEvaluation criteriaEvery case links the thread, the decision and the policy clause appliedIt can be checked on delivered records, not only on a trained model
How much and from whereVolume; historical timeframe; language; source systems50,000 or more cases, 2021 to present, English, Zendesk and ShopifyA supplier can confirm it holds enough history before anything is agreed
Rights and privacyLicensing requirements; de-identification requirementsTraining and evaluation use; customer names, emails and addresses removedPermissions and privacy processing can be checked against a stated use

What success test belongs in a request, and what can it promise?

A success test is a check you can run on the delivered records themselves, such as coverage, linkage and label presence, kept apart from the model result you hope the data will produce. The first can be agreed and checked; the second depends on your own training and evaluation, and no dataset can promise it.

Anthropic's engineering guide to agent evaluations defines a task as a test with defined inputs and success criteria, and says a grader's checks should be clear from the task description so that a result never turns on an ambiguous spec. Hold a data request to the same rule: if two reviewers could read your criterion and disagree about whether a record passes, rewrite it. Format, quality checks and delivery method are agreed for each dataset, so a criterion already in the request gives that agreement a concrete starting point.

The same terms make feedback usable after delivery. This page describes no separate feedback process; the Manage step covers future purchases and the supplier relationship. If a dataset misses the test, or a gap appears once you train, report it against the field or criterion it breaks, for example outcome labels missing for escalated cases in 2022. That turns it into a revised specification for the next request without sharing your model, prompts or evaluation set, and our buyer feedback taxonomy shows how to classify such reports.

Two illustrative requests show the difference between a test and a hope:

  • Usable: resolved refund cases from 2021 onward, each linking the full thread, the order record and the policy clause applied, with the outcome stored as issued, refused or escalated. Pass: at least 95% of delivered cases carry all three links and an outcome. Context, not a condition: the team expects fewer out-of-policy refunds in its offline evaluation.
  • Not scopable as written: customer data that will lift the agent's accuracy by ten points. It names no task, no records and no outcome labels, and its only test is a model result. Rewriting it means naming the failing task, the records that show correct handling and a check on those records.

Why are technical fit and rights fit separate checks?

Because records can match your specification exactly and still not be licensable for your use, while records with clear permissions can still miss the task; evidence for one says nothing about the other. The Assess step on this page names both checks: the data, and the supplier's licensing permissions.

Public datasets show how far the two can drift apart. The Data Provenance Initiative's 2023 audit of widely used AI training data collections reported frequent miscategorization of licenses on popular dataset hosting sites, with license omission above 70% and error rates above 50%. A clean schema said nothing about permission.

Treat an open item on either track as a no until it is closed, and never let a strong result on one track carry the other. Nothing is contracted until a supplier agrees. On the rights track that means the supplier has accepted the scope and terms and put its signature on the agreement before a single record goes to SourceX; on your side, pricing and allowed uses are agreed in a license. The table below sets the two tracks side by side.

Why are technical fit and rights fit separate checks?
Point of reviewTechnical fitRights fit
QuestionDo the records hold the task, context and outcomes you specified?May these records be licensed for the uses you need?
EvidenceA sample prepared like the full set, checked against your evaluation criteriaWhere the records came from and the supplier's licensing permissions, documented before a license is signed
Typical open itemA field, period or outcome label missing from part of the scopeA system, customer contract or record type whose terms are not yet confirmed
Ways to close itNarrow the scope to what the records hold, or look for another sourceDocument the permission, or leave out the records it does not cover
Who decidesYou decide whether a narrowed scope still serves the taskThe supplier decides what it agrees to license; pricing and allowed uses are agreed in a license

What should a dataset sample show before you rely on it?

That it was drawn from the same scope you would license and prepared the same way, with the selection rule, the exclusions and its limits written down. A sample of the best cases shows the best cases, not the dataset.

SourceX's guide to selecting a representative sample says a good sample spans different years, teams and outcomes, failures as well as successes, is prepared exactly like the full set, and is selected by SourceX from data delivered under the supplier's signed agreement. Neither this page nor that guide sets out a statistical sampling protocol, and the candidate suppliers and fit scores in the example request near the top of this page are illustrations, not inventory or validated samples. For a seeded random draw you can run yourself, see our sample representativeness audit.

Representative needs an object. For high-risk AI systems, Article 10(3) of the EU AI Act (Regulation (EU) 2024/1689) requires training, validation and testing data sets to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. Outside that scope the question still helps: representative of what, for which use. Ask for five items with any sample:

Checklist · 5 items

  1. Population: the scope it was drawn from, by system, date range and record type.
  2. Selection rule: random, stratified or hand-picked, and who picked.
  3. Coverage check: the sample's mix of years, outcomes and segments beside the same counts for the full scope.
  4. Exclusions: records removed and why, such as removals during de-identification or records outside the supplier's permissions.
  5. Limits: sample size against the full scope, and the conclusions it cannot support, such as rates for rare outcomes.

0 of 5 ticked. Tick each item as you go.

What should each document in a dataset's diligence packet prove?

One thing each, with its limits stated. This page commits to six kinds of per-dataset evidence, and diligence packets are prepared per dataset because rights, privacy processing and quality differ between datasets. No sample packet is published here, so ask for the packet for the dataset you are evaluating with the Request dataset diligence button above.

Datasheets for Datasets, a 2018 proposal by Timnit Gebru and coauthors, asks that every dataset come with a document covering its motivation, composition, collection process, preprocessing, uses, distribution and maintenance, written for the people who use a dataset as well as those who build it. Its questions are a quick way to spot what a packet leaves out. The table below maps each commitment on this page to what its document should prove and what it cannot.

What should each document in a dataset's diligence packet prove?
Commitment on this pageWhat the document should proveWhat it does not prove
You control the scopeWhich records, systems and date ranges are included, matched to the scope you agreedThat the records are useful for your task
Preparation to specificationThe dataset-specific preparation specification and the checks run against itThat the specification was the right one for your model
Provenance and rights documentationWhere the records came from and the supplier's licensing permissions, documented before a license is signedAnything about quality or coverage
Permitted usesThe allowed uses, duration and restrictions set in the dataset's licensing termsThat a use not named there is allowed
Delivery standardsThe agreed format, quality checks and delivery methodThat no rows were lost before export; ask for counts against the source system
Privacy processingThat personal details were removed, and which processing was appliedThat re-identification is impossible; no method rules it out

Sources for this section

  • Anthropic's engineering guide to agent evaluations defines a task as a test with defined inputs and success criteria and says a grader's checks should be clear from the task description. Source
  • The Data Provenance Initiative's 2023 audit reported frequent miscategorization of licenses on widely used dataset hosting sites, with license omission above 70% and error rates above 50%. Source
  • For high-risk AI systems, Article 10(3) of the EU AI Act requires training, validation and testing data sets to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. Source
  • Datasheets for Datasets (2018) proposes that every dataset come with a document covering its motivation, composition, collection process, preprocessing, uses, distribution and maintenance. Source

Request data

Tell us what your models need.

Share the domain, volume, language, format, timing and licensing requirements.

Request data