Skip to content

Best Practices

Rightmark vs Toloka: a 2026 head-to-head comparison

On this page

Rightmark and Toloka both help AI teams obtain data for training and evaluation. The choice is what you need to receive: original material with a license to use it, expert-built tasks and judgments, or a managed workflow that produces them.

Toloka licenses ready-made expert datasets for evaluation, supervised fine-tuning and reinforcement learning, with commercial-training terms and prices available on contact. It also supplies managed environment and data work. Rightmark sources private code and company records with an ownership pack for the licensing route; on standard licenses, the buyer pays our fee above the owner's agreed payout.

The useful starting point is the work your team cannot already do. Do you need access to a company's original records, or people and engineering to turn a defined problem into training or evaluation data? If you own the material, your decision starts one step earlier: which assets and rights are you willing to license, who will prepare them, and what payment does that grant earn?

Scope a request for the original data you need.

Rightmark sources private code and company records with an ownership pack for the licensing route. Describe your requirements so we can review the scope and next steps.

Request data

Public information checked November 3, 2026. Toloka's product and pricing details come from its own website. Rightmark publishes this comparison.

Rightmark vs Toloka at a glance

Compare the deliverable first, then the price, permissions and work needed to use it.

Rightmark vs Toloka at a glance (2)
DimensionRightmarkToloka
Deliverable and fitOriginal private code or company records with a route-specific ownership pack.Expert tasks and judgments; ready-to-license datasets; managed environment work.
Quality checksRepo Check reports readiness without running tests or builds.Managed automated checks and senior QA; self-service review depends on configuration.
What the buyer pays forA license, with the standard buyer fee above the agreed owner payout.Catalog licensing, managed projects or usage, or self-service contributions and review.
Published price and commitmentScoped quote. Standard fee rate and minimum purchase: not published.Self-service: rates appear in the account or at checkout; inspect the component estimate. Catalog: contact for pricing. Managed rates and minimum: not published.
Existing-asset owner economicsCode pays per license. Archive lump sums and revenue share have different exclusivity terms.Data-partner contact exists; archive-owner payout terms are not published there.
Private material and permitted useScrubbed NDA samples, followed by a license specifying use.Restricted self-service inputs need written preapproval and controls. Inspect the agreement for the selected route.
Approval and controlOwner approves agreement and scope; code can cover later non-exclusive sales.Self-service buyer authorizes the run. Managed scope starts with buyer goals and success criteria.
Data and task coveragePrivate repositories with history, chat, email, tickets and documents.Coding, STEM, enterprise and physical AI catalog data; custom simulated workflows.
Who prepares and runs itOwner-side or buyer-pipeline scans; Rightmark reviews reports and samples and assembles the ownership pack.Self-service buyer defines the task; agent configures it. Managed work covers engineering, collection and QA.
Evidence to examineIllustrative ownership-pack sample. Request evidence for the actual offered material.Supplier-published poolside evaluation case. Request task-specific proof for your offer.
Independent review ratingsNot verified.Not verified.

Check both providers' prices, terms and availability against the proposed scope; neither an estimate nor a worked example is an offer.

How do Rightmark and Toloka compare on licensed training data?

A dataset can be well documented and still be the wrong purchase. Start with what must arrive and the test it must pass, then work through price, permissions, operating responsibilities and proof.

1. Will the delivered data do the job you need?

A repository with years of development history and a task an agent can execute are different deliverables. So are a support archive and a set of expert judgments about support answers. The owner needs the offer to name the material being licensed; the buyer needs it to name what will actually be usable on delivery.

How do Rightmark and Toloka compare on licensed training data? (4)
QuestionRightmarkToloka
What arrives?Original material in graded batches, with an ownership pack for its licensing route.Custom demonstrations, preferences or reasoning tasks with rubrics.
What supports execution?Repo Check reports repository readiness but runs no tests or builds.Managed testbeds include instrumentation and deterministic resets.
How is quality checked?A readiness score does not establish rights or buyer acceptance. Inspect the offered batch.Self-service workflows support LLM checks and contributor consensus; review the selected method.
Who defines acceptable?Put required files, history, links and exclusions into the scope.Custom LLM guidelines and data-point requirements are developed with the client.
What happens to a failure?Correction and replacement terms are still being developed.Submissions below the approved standard are routed for revision.

Toloka can also build custom data from licensed proprietary databases and client materials. This matters when you already have the source and need better tasks or judgments around it. The choice is not settled by whether someone uses the word “licensing.” It is settled by the item in the delivery specification.

A repository that cannot yet run. Imagine Mira, an engineer at Cedar Works, evaluating a private repository with its full git history (an example; the company and the person are invented). The history contains a useful bug fix and its regression test. The build also calls a private service that the buyer cannot access.

Mira can describe two purchases. The first covers the repository and history for an authorized code-data use. The second includes an offline task built from that bug, with a reproducible starting state, replacement for the unavailable service and a test that distinguishes a successful fix. Receiving the first does not complete the second. Someone must scope and pay for that engineering, and the license must permit it.

How do Rightmark and Toloka compare on licensed training data? (8)
Acceptance note for this invented exampleSource-data purchaseRunnable-task purchase
Exact objectRepository version plus agreed historyTask, starting state, dependencies and grading test
Evidence to requestFile manifest, rights record and exclusionsThose records plus build log, execution instructions and test result
Unfinished workRecord the inaccessible serviceAssign who replaces it and proves offline execution
Sign-offNamed reviewer checks the licensed batchNamed engineer runs the task and records pass or failure

The owner should read both specifications too. A promise to provide historical code is narrower than a promise to maintain a working environment. If the buyer needs the latter, decide who takes that obligation before signing.

Bottom line: Toloka is the stronger fit when environment engineering and QA are part of the purchase. Rightmark fits the need for original private material and its licensing pack. Treat additional task construction as its own deliverable.

Ask both: What exactly will we receive, what test makes a unit pass or fail, and who accepts the delivered version?

2. What do buyers pay, and what do asset owners receive?

An owner's agreed payout is only one line in a buyer's budget. A task price pays for a different unit again. Compare the same usable deliverable, including the rights and unfinished work, before deciding whether either proposal is affordable.

How do Rightmark and Toloka compare on licensed training data? (14)
QuestionRightmarkToloka
What is the unit?An agreed license; the standard buyer fee sits above the owner payout.Catalog license; managed project or ongoing usage; self-service expert contributions per item.
What price is published?The standard buyer fee rate and minimum purchase are not published on the inspected buyer and home pages.Self-service rates are shown in the account or at checkout. Catalog pricing is on contact; managed rates are not published.
What else needs budgeting?Name cleanup, export, task construction and acceptance costs.LLM checks are billed on actual use; synthetic generation is billed by input and output tokens.
What happens if work stops?Deal-specific remedies and payment triggers remain to be settled.Purchases are generally final; material breach uncured after five days' written notice permits cancellation, with fees prorated for work performed.
What does an asset owner earn?Code pays per license; archive lump sums are usually exclusive for a term, with non-exclusive revenue share another route.The data-partner contact page publishes no archive-owner payout or royalty schedule.

Rightmark's homepage says owners pay nothing upfront and buyers cover extraction, scrubbing and transfer. Premium and bespoke deals give owners 75% to 85%, agreed in writing. The standard arrangement and the bespoke range answer different questions. Ask which applies to your proposal and which amount the percentage uses. The current extra-charge and payment terms are still being developed, so an owner should obtain those terms for the actual deal.

One illustrative self-service calculation. Assume 1,000 accepted contributions from one expert each, a USD 0.50 payout per contribution and a 33% provider commission. Expert payout totals USD 500; the assumed commission adds USD 165, giving USD 665 before any separately priced preparation or review. All three inputs, including the commission, are invented assumptions for this calculation. This is neither a Toloka quote nor a return to an archive owner. Adding a second independent answer per item would also change the work purchased, so keep “items” and “expert contributions” separate in your estimate.

Use the same sheet for both offers. Leave a field blank until someone supplies the amount and its basis; a blank is a follow-up, never a zero.

How do Rightmark and Toloka compare on licensed training data? (18)
Offer worksheetRightmark proposalToloka proposal
Deliverable, version and quantity____________________
Permitted uses, term and exclusivity____________________
Owner or license amount, and its unit____________________
Provider fee, rate and calculation base____________________
Preparation and task-building charges____________________
QA, acceptance and expected rework____________________
Minimum commitment and cancellation charges____________________
Buyer's internal work, tax and transfer costs____________________
Original owner's payment amount and trigger____________________
Buyer total for this complete scope____________________

For a buyer, total the license or production price, provider fee, preparation and verification work, then add the work your own staff must do. For an owner, read the proceeds line alongside exclusivity, retained uses and the payment event. A larger payment tied to a broader grant or later milestone is a different offer, even if the headline percentage looks attractive.

Bottom line: Neither the self-service commission nor Rightmark's owner share establishes the cheaper complete purchase or the higher owner return. Choose between written offers for matching rights and deliverables.

Ask both: What is the full charge for this deliverable, including preparation and rejected work, and which amount and milestone govern the original owner's payment?

3. Whose information moves, and what may the buyer do with it?

A buyer needs to know whether the planned test is allowed before uploading a sample into a model or handing it to another provider. The owner needs to know how far that permission travels. Put the intended recipients, systems, uses and retention into the request while the material is still under your control.

How do Rightmark and Toloka compare on licensed training data? (24)
QuestionRightmarkToloka
Underlying rightsOwners retain ownership and grant specified uses.Paid self-service rights are summarized below; request the separate catalog or managed agreement.
Before licensingRepresentative scrubbed samples under NDA precede the signed use, scope and term.Catalog samples can be requested. Confirm the allowed test with that request.
Additional usesSome code routes allow partner sublicensing, own use or derived datasets, disclosed before signing.Check the provider's model-use permission and opt-out scope below.
Transfer methodAdmin-approved revocable access or exports.Managed environments use client-approved handling and controlled testbed credentials.

Toloka's self-service agreement requires written preapproval and additional controls for Restricted Content, including confidential business information. After full payment, customers own custom Output, subject to Seller Background IP. Sellers retain pre-existing or independently developed IP; when embedded and necessary to use Output, buyers receive a non-exclusive, worldwide, perpetual, royalty-free license solely within that Output. Customers also permit Toloka to train or improve its internal models using customer Data and Output. Written opt-out takes effect prospectively on receipt and excludes Toloka Data: de-identified and/or aggregated data, derivatives, insights, metrics and telemetry derived from services, customer Data, Output and service usage. Trial Output instead remains Toloka-owned, with a revocable internal-evaluation license limited to the trial. These are self-service terms, not a summary of a catalog or managed-service contract.

For the invented Cedar Works repository, put three proposed actions on separate lines: inspect the files, train a model on them, and let another company build tasks from them. Name the recipient and the systems for each. The owner may be comfortable with the first and require different terms for the others. A confidentiality agreement alone does not answer all three questions. For a specific Rightmark sample, the permission for the proposed model test still needs confirmation.

The same discipline helps an archive owner avoid accidental overreach. If a licensed collection excludes customer records, the sampling and cleanup plan must preserve that exclusion. A useful technical export should not quietly widen the contractual scope. Record which fields and systems were removed so a buyer can judge the remaining material accurately.

Bottom line: Choose the offer that explicitly covers your intended use and preserves the rights you need to retain. Neither a source-data license nor an output-ownership clause is a shortcut around checking inputs, third-party material and downstream recipients.

Ask both: Which signed permission covers these exact inputs, model tests, derivatives and recipients, and what changes or ends when we withdraw, opt out or reach the term end?

4. What can buyers and owners approve or change?

An owner approves a grant of rights; a buyer approves spending and delivery requirements. Those decisions may belong to different people. Make the approval sequence visible before anyone interprets an early yes as authority to start a later step.

How do Rightmark and Toloka compare on licensed training data? (32)
DecisionRightmarkToloka
Owner agreementOwner approval covers agreement and scope; code signing can authorize future non-exclusive sales.Existing-asset supplier approvals are not defined on the partner contact page.
Buyer specificationWritten scope names families, filters, formats, de-identification, volume, exclusivity and cadence.Self-service setup exposes data structure, task UI, guidance, pricing and QA for review.
Starting paid workName the spending and acceptance signers in the offer.The buyer launches the run; the agent configures it.
Later changesAgree how a changed source, permission or acceptance test amends the scope.The Pipeline API can copy a designed pipeline for another run; confirm how later edits affect that copy.
Quality requirementsName who can approve samples and resolve a disputed test.Buyers can revise the agent's quality criteria.

Toloka's managed environment work starts with buyer goals, constraints and success criteria. That is a useful point to specify who can accept a change in difficulty, a new workflow or a different grading rule. A purchasing contact may approve a budget increase while only the engineering lead can decide whether a task still tests the intended behavior.

For an owner using Rightmark, future licensing authority deserves the same attention as the first sale. Decide whether the agreed scope includes later repository versions, new records or only the named collection. If a future sale is already authorized, the practical control is the agreement you sign now. Keep a copy of the exclusions and withdrawal conditions with it.

For the buyer, an approval log can be small: decision, approver, version, date and what changed. If the source scope changes after a sample succeeds, rerun the relevant acceptance checks against the new version. A passed test on an earlier collection should not silently accept its replacement.

Bottom line: Rightmark's published approval concerns the license and collection scope. Toloka self-service makes the buyer responsible for task knowledge and launch approval. Match each decision to the person who can actually make it.

Ask both: Who approves the source scope, sample tests, spending and acceptance, and which later changes require a fresh approval?

5. Do the sources, tasks and locations match the use case?

A company's working history can reveal how decisions unfolded. A purpose-built task can isolate a behavior and give it a repeatable test. Neither substitutes automatically for the other, and a category label such as “enterprise data” leaves that distinction unresolved.

How do Rightmark and Toloka compare on licensed training data? (40)
RequirementRightmarkToloka
Original workplace contextLinked records can connect a thread, its ticket and the code that fixed it.Custom work can use client material and licensed databases.
Designed task worldsDefine any task-building work separately from the source license.Managed work includes enterprise-tool replicas and simulated companies.
Code coveragePrivate repositories must never have been public, forks or published packages.Terminal and scientific-coding datasets are listed; confirm which version is ready to license.
Domain and languagePut both into the request and inspect the relevant sample.Custom post-training lists computer science, law and finance, with English, German and French among its languages.
Location and handlingIdentify where records originate and which people and rights they involve.Review the node's audience filters and contributor guidelines before launch.
Current availabilityThe buyer page's pooled lots remain in preparation.Confirm availability and delivery status for the exact named dataset.

Toloka's catalog spans enterprise tasks, coding, STEM and physical AI data. Ask for the named dataset and its present delivery status. A live sample from one collection does not establish that a different collection can ship on the same date.

For original workplace records, try following one permitted chain from request to resolution. Does the ticket retain the link to the relevant discussion? Does the code reference survive export? If the buyer needs the reasoning behind a change, an export of final documents alone may miss it. Owners can make this review easier by listing available systems, years and broken links before a price is discussed.

The geographic question is equally concrete. State the source languages, the countries relevant to the records, and the intended processing locations. A provider's contributor-selection settings do not supply permission to process restricted inputs. Rightmark does not currently take health records, including patient, medical-imaging or veterinary records.

Bottom line: Buy historical records when the source history is essential; buy a designed task when controlled execution or expert judgment is the requirement. Confirm the actual collection, language and available components before treating either route as a match.

Ask both: Can you show a permitted sample from the required source, domain, language and period, and identify which requested components are available now?

6. Who performs the preparation, QA and delivery work?

A promising sample can leave a large amount of work between approval and useful delivery. Owners may still need to export or clear records, while buyers may need to assemble tasks and check them. Name the person and cost for each handoff rather than leaving “preparation included” undefined.

How do Rightmark and Toloka compare on licensed training data? (48)
WorkRightmarkToloka
Define requirementsBuyer sends a specification; Rightmark replies with sourcing scope, volume and timing.Self-service buyer supplies task and domain knowledge.
Prepare existing materialScans run with the owner or buyer pipeline; Rightmark reviews reports and scrubbed samples.Custom LLM services include preparation, annotation and enrichment of client data.
Build tasks and environmentsName who fixes dependencies and proves execution.Managed RL work covers engineering, collection and QA.
Operate the workflowRightmark assembles the route's ownership pack. Confirm non-code cleanup responsibility.Self-service uses an agent-configured pipeline; managed services are separately offered.
Inspect and accept deliveryAssign a buyer acceptance owner and test.Self-service results are inspectable during a run and exportable when complete.

For an archive, split the preparation line into export, sensitive-field removal, third-party-rights review, link preservation and packaging. Put one responsible party beside each. If an owner can export tickets but cannot export their attachments, record the gap and let the buyer decide whether the remaining collection still serves the task. Do not let a volume count hide a missing component.

For code, return to Mira's inaccessible service. The owner might be able to supply setup documentation but lack permission to transfer the service itself. The buyer might have engineers who can replace it, or want that work commissioned. All three facts affect the offer: what may transfer, who completes the environment and who accepts the result.

A useful handover contains the received version, known exclusions, outstanding preparation, the assigned owner of that work and the next acceptance test. That gives both teams a way to discuss a delay or scope change without arguing about whether a broad word such as “ready” was satisfied.

Bottom line: Toloka managed services fit a buyer purchasing engineering, collection and QA together. Rightmark fits sourcing and licensing original material. Use the scope to assign every remaining preparation and acceptance step.

Ask both: Who is responsible for each preparation and acceptance step, what do our people still have to do, and which unfinished work changes the price or delivery date?

7. What evidence supports fit for this particular purchase?

Look for evidence that reaches the same decision your team must make. A specimen ownership folder helps counsel know what to request. A completed evaluation workflow helps an engineering team understand how judgments were made. Neither alone tells an archive owner what their own license will earn.

How do Rightmark and Toloka compare on licensed training data? (56)
EvidenceRightmarkToloka
Public exampleThe ownership-pack sample is illustrative, not from a real dataset.The poolside case describes developer-intent-based pairwise evaluation.
Ongoing quality workflowRequest a relevant delivered batch and its acceptance record.The Shopify case describes expert corrections and sampled Shopify review.
Independent ratingsNot verified.Not verified.
Evidence for your offerPermitted sample, ownership documents and the agreed acceptance test.Task sample, rubric and review or execution evidence for the selected route.

In its October 15, 2025 poolside case, Toloka describes experts comparing two developer answers after identifying the user's intent. The evaluator chooses a response and explains why. A technically plausible answer can lose because it does not help with the actual task. That is a useful procurement question: will the supplier deliver the judgment alone, or the reason that lets your team investigate disagreement?

Toloka's August 18, 2026 Shopify case describes de-identified production conversations routed for expert correction, versioned annotation guidelines and Shopify's review of sampled labels. It illustrates a division of responsibility worth discussing with either provider: who supplies the examples, who interprets them and who resolves a disputed judgment? These accounts were published by Toloka.

In Ksenia Se's June 14, 2025 Turing Post interview, Toloka CEO Olga Megorskaya says, “The human expert is the benchmark.” Her perspective as Toloka's leader points to a practical issue: deciding what a correct answer means for your task. The name of the expert profession is only a start; ask to see the rubric and how ambiguous cases are resolved.

Bottom line: Toloka's named evaluation cases provide a stronger public account of expert-workflow delivery. For either offer, request proof about the version, task and rights you intend to buy. A relevant sample and acceptance record are more useful than an unrelated success story.

Ask both: Which comparable deliverable can you demonstrate with permission, and what evidence links its scope, acceptance criteria and payment terms to our proposed purchase?

What Rightmark does differently among Toloka alternatives

These are concrete reasons to consider Rightmark when original material is the missing input:

  • An ownership pack alongside the source material. Our buyer offer pairs private code and company records with an ownership pack for the licensing route. Request the documents that support the specific collection, then inspect its known gaps. The pack and the technical sample answer different parts of the same purchase.
  • A code report before repository transfer. Repo Check runs locally and produces share.json, a summary of counts and fixed labels the owner chooses whether to share, without source-code lines. Use it to identify the next technical questions. It does not establish ownership or acceptance, and it does not run the repository's tests.
  • Licensing choices that preserve ownership. The owner keeps ownership under the licensing model. Code is non-exclusive by default; archive lump sums are usually exclusive for a term, with non-exclusive revenue share an alternative. Compare the rights you retain over the same period, including any authority for later code sales.
  • Separate owner payout and buyer fee. On standard licenses, Rightmark deducts nothing from the agreed owner payout and charges the buyer its fee on top. Put both lines in the worksheet. Then identify preparation charges and the owner's payment trigger for the actual proposal.
  • Written scope before samples and the license. The buyer sequence sets scope, then provides reports and scrubbed samples under NDA, before signing. Specify the proposed sample test and its allowed systems and recipients at that stage. Exact model-test permission remains deal-specific.

When is Toloka the better choice?

Choose Toloka when demonstrations, preference judgments or reasoning tasks with rubrics are what your team needs delivered. Its managed environment service is the better fit when you also need engineering, collection and QA supplied together. A team trying to evaluate an agent's decisions needs someone to define correct behavior and build evidence around it; access to source material alone will not finish that job.

Toloka also offers preparation and enrichment of existing client data. If you already hold the necessary material and permissions, discuss the work still missing. A catalog license is another route when a ready-made set fits; self-service suits a team that wants an agent to configure its labeling workflow. Decide how much task definition and acceptance your people will own before choosing the service level.

For an archive owner approached through the data-partner route, ask for the actual licensing proposal. The offer should identify the source assets, rights, preparation responsibilities and proceeds. Contributor pay for doing tasks answers a different economic question.

When is Rightmark the better choice?

Choose Rightmark when your missing input is original private code or company records with an ownership pack and a defined license. The fit is strongest when you can describe why that original history matters and test a permitted sample against the requirement. For a repository, state whether you need source and history or an accepted runnable task. For an archive, name the systems and relationships that must survive export.

For owners, this is a route to consider when you want to license existing assets while keeping ownership and can accept the proposed scope. A non-exclusive code license, a term-exclusive archive lump sum and non-exclusive revenue share create different choices. Read future licensing authority, retained uses and the payment trigger together. A favorable payout line does not compensate for rights you did not intend to grant.

Previously public code and forks fall outside the current code eligibility rule. Health-data intake is not open. For eligible material, the next decision is whether a specific buyer's requirement, preparation plan and proposed rights fit the collection you actually own.

Consider a third option if you need original assets and expert-built AI data

Some purchases need both. You can license source material and commission task construction under separate scopes, or use already-authorized internal material with your own engineering and expert team. Separating the work makes its costs visible, but also leaves you responsible for the handoff.

Before combining providers, make sure the source permission covers the second provider's work, identify who removes excluded material and agree which deliverable the final acceptance test measures. Put the two budgets together, including your team's integration time. Otherwise, each supplier can complete its own brief while the finished dataset still misses yours.

An owner with unresolved consent or export rights should first isolate the material they can authorize. A buyer who cannot describe a passing result should first write that test. Either decision can save an unnecessary procurement cycle.

Relevant reading: Rightmark vs. Scale AI and code datasets vs. RL environments.

The decision is still yours

Choose the thing you need to obtain: original material, expert-built tasks and judgments, or both. Then take the Ask both questions into a written offer covering the delivered version, permitted uses, remaining work, acceptance and payment.

For buyers, compare the total cost of making that deliverable usable. For owners, compare the proceeds with the exact rights granted and the milestone that releases payment. Keeping those two views on the same page makes the decision easier to defend on either side of the agreement.

See where your existing data could fit.

Answer a few questions about your data to see an indicative range and possible licensing routes. The free appraisal requires no files or system access and gives an estimate, not an offer.

Appraise your data

Sources

Source dates distinguish publication, publisher updates and archive captures.

What brings you to Rightmark?

License your data, find data for your team, or introduce an owner.

What brings you to Rightmark?