Skip to content
Synthium Labs
Start a conversation

Data assets and expert intelligence
for frontier AI.

From acquiring valuable real-world data to transforming complex information into expert-curated training and evaluation assets, Synthium helps AI teams build on better data.

See how it works

Path one

Acquire data

For teams that need external data for training, post-training or evaluation, and for organizations that hold data they could license.

  • Who it is forTeams seeking data, and organizations holding data they may be able to license.
  • You receiveAn assessment of fit, rights and provenance first. Where it proceeds, an agreement that defines permitted use, and the documentation behind the data.
  • Not assumedThat a given dataset exists, can be licensed, or suits your intended use. That is established per engagement.
Discuss acquiring dataLicense or sell data to Synthium

Path two

Prepare and improve data

For teams that already have data or a task and need it turned into expert-reviewed training or evaluation assets.

  • Who it is forAI labs and model teams with data, or a task that needs domain judgment.
  • You receiveAnnotated, reviewed records against guidelines you have agreed, with the documentation of how they were made.
  • Not assumedAny task type, scale or turnaround. Scope is confirmed before work starts.
Discuss a data curation projectSee what curation involves

Two paths, compared

Acquire data, or prepare and improve it.

They answer different questions and can be combined. Most teams start with one.

Synthium Data: acquiring data compared with preparing and improving data
AspectAcquire dataPrepare and improve data
The questionWhere can the data I need come from, and may I use it for this purpose?How do I turn the data or task I have into reliable training or evaluation assets?
You bringThe need in words (what, for what use), or a description of data you hold and may license.The data in words, or the task, plus the standard it has to meet.
We doAssess relevance, rights and provenance before any commercial discussion. Source or license where it is a fit.Define guidelines with you, qualify and calibrate experts, annotate, review, adjudicate and validate.
You receiveAn assessment, then where it proceeds, an agreement defining permitted use and the data with its documentation.Expert-reviewed records with the guidelines, review decisions and adjudication records behind them.
Not assumedThat a dataset exists, can be licensed, is exclusive, or suits your intended use.Any task type, scale or turnaround. These are confirmed first.

Synthium Data

Two capabilities. One data infrastructure.

Synthium works on both sides of the data problem: where valuable data comes from, and what it takes to make it good enough to learn from.

Capability one

Data Acquisition & Licensing

We work with organizations that own valuable data assets to assess, acquire, or license eligible datasets for AI development.

Who it serves
Organizations that own or control valuable data assets and have the rights to offer them.
The problem
Valuable operating records often sit unused in systems a company already runs, with no clear route to responsible AI use.
What Synthium does
Assesses relevance, rights and provenance first. Acquisition is selective: nothing enters the pipeline by default.
What you receive
An honest assessment and, where it is a fit, an agreement that defines permitted use. Terms are not published, and nothing moves before an agreement exists.
License or sell data to Synthium

Start here. Which of these does your company use?

Chat, email and files
CRM, support and calls
Engineering, finance and data

A short form, no commitment, and nothing is uploaded: you describe your data in words. Choosing a tool tells us what you run, not that a connector exists.

Three steps to a check.

  1. Check your data

    Tell us which systems your company runs, then describe what you hold in words. No files, no samples, no credentials.

  2. Review rights and provenance

    We assess ownership, authority and provenance before any commercial conversation. Most opportunities will not proceed.

  3. Agree before anything moves

    If it is a fit, permitted use is defined by agreement and a secure transfer route is agreed. Nothing moves by default.

ILLUSTRATIVE EXAMPLE

EXAMPLE-DATASETNo review completed
DomainOpportunity
Software engineering
SourceOpportunity
Organization-held records
Ownership and authorityRights review
Not assessed
ProvenanceQuality review
Not assessed
Permitted useAgreement
Defined by agreement, not by default
Expert validationCuration
Not started
VersionDelivery
Not issued
A record shows what has been checked and what has not. It does not mean a dataset is held, available, or for sale.

Capability two

Expert-Led Data Curation & Annotation

We work with AI teams to transform raw data and complex domain-specific tasks into structured, expert-reviewed datasets for training and evaluating AI systems. That includes data annotation and labeling where the task calls for it.

Who it serves
AI labs and teams that need annotation, expert judgments, dataset preparation or quality assurance.
The problem
Specialised tasks need domain judgment a generic annotation workforce cannot supply, and a dataset is only as trustworthy as the process behind it.
What Synthium does
Defines the task and guidelines with you, qualifies and calibrates domain experts, produces structured annotations and judgments, then audits quality and adjudicates disagreements.
What you receive
Expert-reviewed datasets, with the guidelines, adjudication records and quality report behind them where the engagement includes them.

We confirm what is in scope, and at what volume, before an engagement starts. It is never assumed from this page.

Discuss a data curation project

ILLUSTRATIVE EXAMPLE

Preference comparison

Task

A customer asks for a refund outside the stated refund window. Which reply handles the request better while following the policy?

Response A

Of course, we'll refund the full amount today. Sorry for the trouble!

Response BPreferred

I can't refund this order under our 30-day policy, but I've arranged a store credit for the same value and explained how to use it.

Criteria

Each criterion, judged for Response A and Response B
CriterionAB
Follows the stated policyDoes not meetMeets
Answers the question askedMeetsMeets
Appropriate toneMeetsMeets

Expert rationale

A promises a refund the policy does not allow. B reaches a fair outcome without contradicting the policy, and says so plainly.

Two reviewers differed on tone. An adjudicator decided and recorded why.

How expert-led data work happens

From a defined task to a documented dataset.

The route adapts to the engagement: the task type, the expertise it needs and the intended use decide which steps matter most. These are the steps most engagements share.

  1. Define the taskIntended use, task type and quality requirements
  2. Guidelines and rubricsTask-specific instructions and evaluation criteria
  3. Qualify expertsContributors and subject-matter experts selected for the task
  4. CalibrateAnnotators aligned against representative examples
  5. Annotate and judgeStructured annotations, rankings and written rationales
  6. Audit and adjudicateQuality checks, and a qualified decision where reviewers differ
  7. Validate and deliverThe dataset, with the documentation behind it

What we assess

What we produce.

The kinds of work, where the engagement calls for them. Each is built from expert judgment and checked before it is delivered.

Structured annotations and classifications

Labels, categories and extracted fields applied against task-specific guidelines.

Expert-reviewed question-and-answer pairs

Domain questions with answers written or verified by a qualified expert.

Preference comparisons and rankings

Side-by-side judgments between responses, with the reasoning behind each choice.

Rubric-based evaluations

Responses scored against criteria agreed before scoring begins.

Critiques with supporting rationale

Written feedback on model output that explains what is wrong and why.

Cleaned and enriched datasets

Records normalized, deduplicated and enriched, with consistency checks applied.

Expert-written examples and corrections

Worked examples, or corrected versions of model output, written by a domain expert.

What you receive

What the engagement hands over, and how firmly.

The documents and artifacts around the data. They are what let you trust it, check it and use it within the terms agreed.

Acquire data

  • Fit, rights and provenance assessment

    Whether the data is relevant, who owns or controls it, whether they have the authority to license it, and where it came from. Done before any commercial discussion.

    Standard

  • Permitted-use terms

    An agreement that states what the data may be used for, and what it may not. Licensing a dataset does not mean Synthium owns it or may use it for every purpose.

    Standard

  • Dataset manifest, schema and field definitions

    What is in the delivery, how it is structured, and what each field means.

    Standard

  • Known limitations and permitted-use notes

    Gaps, biases and restrictions we know about, stated rather than left to discover.

    Standard

  • Coverage and suitability assessment

    How well the data covers the use you intend, and where it does not.

    Optional

  • Quality validation summary

    The checks run on the delivered data and what they found.

    Optional

  • Delivery route and format

    How the data is transferred and in what form. Agreed securely once an agreement exists; nothing moves before.

    Scoped per engagement

Prepare and improve data

  • Annotation guidelines and rubrics

    The task-specific instructions and criteria the work is done against, versioned so a record can be traced to the version that governed it.

    Standard

  • Annotated, reviewed records

    The labels, judgments, rankings or written outputs, in the schema you specify.

    Standard

  • Reviewer decisions

    Each independent reviewer's call on an item, kept so a disagreement can be seen.

    Optional

  • Adjudication records

    Where reviewers differed: who decided, what they decided and why.

    Optional

  • Quality summary

    Sampling and audit results for the delivery, reported plainly. We do not publish headline figures; real numbers belong to an engagement.

    Optional

  • Correction log and provenance

    What was corrected after review, traceable to the item and the reviewer.

    Optional

  • Volume, turnaround and task type

    Not assumed from this page. Confirmed before work starts, and written into the scope.

    Scoped per engagement

Which of these an engagement includes is agreed before work starts. A listed item is a capability, not a promise that every engagement contains it.

A sample record

One item, from source record to adjudicated decision.

This is the shape of what expert-led curation produces for a single item: the source, the instruction, two independent expert outputs, where they differed, and how the difference was settled.

ILLUSTRATIVE EXAMPLE

Sample record: classification with review and adjudication

Source record (synthetic)

record_id:  EXAMPLE-1042
channel:    email
text:       "I was charged twice for the March plan and the
             second charge is still pending. Please fix this
             today or I will cancel."

Task and instruction

Assign one category and one urgency, using the guideline definitions. If the message names a deadline or a threat to cancel, urgency is high. Give a one-line reason.

Expert outputs (two independent reviewers)

ReviewerCategoryUrgencyReason given
Reviewer 1Billing: duplicate chargeHighThe message threatens cancellation.
Reviewer 2Billing: duplicate chargeNormalThe second charge is pending, not settled.

Review decision

The reviewers agree on the category and differ on urgency. The item is escalated; it is not resolved by averaging.

Adjudication

Decision: urgency is high. Guideline version 1.2, section 3: a stated cancellation threat sets urgency to high whatever the charge status. The guideline is clarified to say that a pending charge does not lower urgency, and Reviewer 2 is shown the reasoning.

Provenance and handling

  • Source: synthetic text written for this example.
  • Rights: none to establish; no customer data is involved.
  • Guideline version: Example v1.2.
  • Correction log: one label corrected, traceable to its reviewer.
A synthetic record written for this page. It is not a customer record or a delivered dataset, and it carries no accuracy, agreement or throughput figure. Which checks apply to an engagement (second review, audit sampling, adjudication) is set per engagement.

How it runs

Quality is the difference.

These are the operating principles behind every dataset. They are why expert-led data work differs from commodity annotation, and none of them is a number we are claiming.

  1. Qualification before production

    Contributors are assessed against the expertise a task actually needs before they touch production work.

  2. Task-specific guidelines and calibration

    Clear instructions per task, and annotators calibrated against representative examples before scaling.

  3. Expert review and adjudication

    Specialized and ambiguous items go to domain experts. Disagreements are adjudicated and explained, not averaged away.

  4. Independent checks, traceable corrections

    Quality audits are independent of the people who did the work, and every correction can be traced.

  5. Rights, confidentiality and handling

    Appropriate data rights, confidentiality and handling controls apply to the data and to the people who work on it.

Who does the work

Curated by people who understand the domain.

Data work is only valuable when someone who understands the domain judges it. We staff each project with experts qualified in the relevant field, and confirm who is needed before work starts.

  • Software engineering workflows
  • Customer support operations
  • Financial operations and compliance
  • Legal and policy processes
  • Healthcare and clinical workflows
  • Sales and business development
  • IT operations and incident response

Engagement and fit

Who this is for, and what to bring.

A first conversation needs a description, not data. Bring what you can say in words.

What to bring to a first conversation

  • The intended useTraining, post-training or evaluation, and the capability you are trying to improve.
  • The data or task, in wordsWhat it contains, where it came from, what period it covers. No samples or files.
  • Your standardHow you would judge a good result, and any constraints on handling.

A good fit when

  • You need data you may legitimately use, and you want its rights and provenance made explicit.
  • You have data or a task that needs domain judgment rather than generic labelling.
  • You can describe the outcome you need and who should judge it.

Not the right fit when

  • You want a guaranteed dataset, volume or price before any assessment.
  • You expect the data itself to be sent through a web form or by email.
  • The task needs no domain judgment and no review.

Describe what you have or need in words. Please do not send datasets, samples, personal records or credentials through either form. If we proceed, we agree a secure route first.

Both capabilities

Start with the data, the work, or both.

Some engagements begin with a company's existing data assets and continue through curation, annotation, enrichment and validation. The stages are configurable: you can engage Synthium for acquisition or licensing, for curation and annotation, or for both.

  • Acquisition or licensing

    You hold data and want to know whether, and how, it could be used for AI. Rights and provenance come first.

  • Curation and annotation

    You bring the data or the task. We prepare it: guidelines, expert judgments, quality assurance, delivery.

  • Both, end to end

    We assess a dataset, then prepare the approved assets for an agreed training or evaluation use.

An example of the combined path

An organization licenses an eligible proprietary dataset. Synthium assesses its suitability, organizes and curates the data, applies expert annotation where required, performs quality assurance, and prepares the approved assets for an agreed AI training or evaluation use.

Licensing a dataset does not mean Synthium owns it outright, may redistribute it, or may use it for every purpose. Permitted use is defined by agreement, dataset by dataset.

Choosing between offerings

Which Synthium offering fits the problem?

They are separate offerings, and they work together where a project needs more than one. Most teams start with one.

When to choose each Synthium offering, and what each delivers
OfferingChoose it whenYou receive
Synthium Data (this page)You need data acquired or licensed, or you have data or a task that needs expert curation and annotation.An assessment of fit and rights, or an expert-reviewed dataset, with the documentation behind it. Scope is agreed first.
Synthium TrainYou are improving an agent or model and need RL environments, or expert-generated post-training data.Environments with tasks and verification you run in your own training stack, and/or datasets in your schema.
Synthium AgentsYou need to know how well a model or agent performs on realistic tasks, and exactly where it fails.An evaluation specification, traces, expert-reviewed findings and a failure analysis you can act on.
Synthium Red TeamYou want authorized adversarial testing of an AI system and documented, reproducible failures.A scoped test plan, findings with evidence and a severity rationale, and mitigation recommendations.

Keep exploring

Related solutions

Also see:Security and data handlingThe quality system

Questions

Questions about Synthium Data

Can you get me any dataset I name?
No. Whether a dataset exists, can be licensed, is exclusive and suits your intended use is established for each engagement, and some requests cannot be met. We say so after the assessment rather than before it.
Does licensing a dataset mean Synthium owns it?
No. A licence does not give Synthium ownership, the right to redistribute, or the right to use the data for every purpose. Permitted use is defined by agreement, dataset by dataset.
Should I send you sample data?
No. Describe it in words. Neither form accepts files, and we ask you not to paste samples, personal records or credentials. If we proceed, we agree a secure route first.
How do you handle experts who disagree?
Disagreement is investigated and adjudicated by a qualified reviewer, and the decision and reasoning are recorded. It is not averaged away. Whether the records are part of a delivery is set per engagement.
What accuracy or agreement can I expect?
We do not publish figures. Quality is described by the method (qualification, calibration, review, adjudication, audit) and evidenced by the records of an engagement, shared under NDA or in a pilot.
What if I need both acquisition and curation?
That is a common shape. We assess a dataset, then prepare the approved assets for the agreed use. See the combined path above, and say so in the first conversation.

Next step

Bring us the data, the task, or both.

Two ways to start. Describe what you have or what you need, and we will tell you honestly whether and how we can help.

Acquire or license data

Two ways in. If you want data sourced or licensed for a training or evaluation need, describe the need. If you hold data and want to offer it, tell us what it is. Either way, rights and permitted use are assessed before anything is discussed commercially.

Discuss acquiring dataLicense or sell data to Synthium

Prepare and improve data

For AI labs and teams that have data or a task and need annotation, expert judgments, dataset preparation or quality assurance, for training or for evaluation. Bring the task, the data in words, and the standard it has to meet.

Discuss a data curation project

Describe what you have or need in words. Please don't send datasets, samples, personal records or credentials through either form. If we proceed, we agree a secure route first.