Skip to content

Best Practices

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

On this page

Innodata and Rightmark both help AI teams obtain data for model training and evaluation. The decision is what you need supplied: access to particular source material, prepared examples for a defined task, or the expert work of designing and scoring an evaluation.

Innodata offers prebuilt question-and-answer and question-solution-answer datasets alongside expert-created fine-tuning data. Neither offer page publishes prices or billing units. Rightmark sources private material to a buyer's specification and, on standard licenses, charges the buyer its fee on top of the owner's payout.

The distinction can be subtle. A publisher's question bank may already contain worked solutions, while a buyer may still need answers checked against a different task or a new scoring rule. For the owner, that is a choice between licensing an existing asset and taking on additional work. For the buyer, it determines what must arrive before the team can use the data.

Need source material for your model task?

Rightmark sources private material to your specification and checks title and consent. Tell the team what you need, and it will reply with what it can source, volumes and timing.

Request data

Rightmark publishes this comparison, based on public information checked in November 2026; Innodata's product and commercial details come from its own website.

Rightmark vs Innodata at a glance

Start with the deliverable each proposal includes, then compare fees, permissions and the work left to your team.

Rightmark vs Innodata at a glance (2)
CompareRightmarkInnodata
Best fit and deliverableParticular proprietary material with its rights documented; a question bank may already contain worked solutions.Prepared task data or a commissioned evaluation.
Buyer price and commitmentStandard-license buyer fee on top of owner payout. Matching total price, fee amount and minimum commitment: not published.Catalog and managed-work prices, billing units and minimum commitments: not published.
Owner paymentStandard licenses have no fee deduction from owner payout. Premium/bespoke owner share: 75% to 85%, agreed in writing.Supplier and licensing inquiries contemplated; standard owner payment terms not published.
Permitted use and handlingSeparate document training, retrieval and display rights. Samples are for evaluation only under NDA; agree the particular test.A universal dataset grant is not published. Obtain the requested license and sample permissions.
Approval and withdrawalCode signing covers future non-exclusive sales; withdrawal applies to code not yet licensed. Other routes need their own terms.Asset-specific approval and withdrawal terms: not published in the inquiry policy.
Coverage and task fitPrivate code, archives and documents; match existing structure to the intended task.Prepared coding datasets; expert work across languages and formats.
Preparation and owner workArchive extraction, scrubbing, transfer and fees agreed in writing. Confirm additional non-code preparation and task checking in the proposal.Question generation and answer verification, scenarios and rubrics: establish which services are included.
Availability and timingSourceability, volumes and sample timing confirmed for the request. Lots labeled in preparation are unavailable.Confirm the selected deliverable, requested changes and delivery date.
Relevant proofIllustrative ownership packs and records of actual checks before buyer review.Anonymous evaluation case describing joint refinement of rating guidance.

Compare fees, terms and availability against the matching written proposals. "Not published" refers to the relevant public pages reviewed.

How do Rightmark and Innodata compare for a defined model task?

A large archive can leave your team with substantial task-building work, while prepared examples can still miss the behavior you need to test. Start with usefulness, then examine cost, permissions, control, coverage and workload before deciding what the evidence proves.

1. Will the delivered data support the task you actually need?

The expensive surprise is receiving the agreed files and still having no way to judge an answer. A source collection, a set of training examples and a scored model evaluation solve different problems. Owners need the same distinction: permission to use an existing answer does not define an obligation to write or check another one.

How do Rightmark and Innodata compare for a defined model task? (4)
QuestionRightmarkInnodata
What arrives?A scoped source asset, such as an archive or question bank; identify its counted unit.Expert-created task data is offered. Specify the deliverable separately from any evaluation service.
Are answers included?Document collections can retain worked solutions, rubrics and graded answers. Agree any new authoring or verification separately.Answer verification and question generation are listed fine-tuning tasks. Confirm the purchased work.
How is an answer scored?An existing rubric can travel with its source. Agree a model-scoring procedure separately.Benchmarks use expert-authored scenarios and scenario-specific rubrics.
What can you inspect first?Representative scrubbed samples under NDA, for evaluation only; the exact model test needs permission.The catalog invites sample requests. Establish the allowed test before using one.
What counts as accepted?Name the checker and task test; standard acceptance and remedy terms are developing.Acceptance and correction terms are not published on the benchmark offer page. Put them in the proposal.

In a January 2025 public discussion of post-training, Nathan Lambert describes beginning with the behaviors and evaluations a team cares about, then obtaining or creating queries that exercise them. For procurement, that means writing the test question before asking for a dataset count. If the job is diagnosis, decide whether a correct final answer is enough or whether the reasoning must identify the fault, rule out plausible alternatives and justify the next check.

An invented example: A technical-training publisher owns a question bank with worked answers. An AI team wants fault-diagnosis tasks with reasoning it can check. The publisher and AI team in this example are invented; this is a purchasing decision, not a reported transaction.

Suppose a source question gives a symptom and asks for the most likely fault. Its worked answer names the fault, but the buyer also wants the model to select a diagnostic test and explain what each possible test result would mean. The original question remains useful. It simply does not contain every component of the buyer's task.

There are two sensible proposals. One licenses that particular bank and names who will check the existing answers and create the missing scoring guidance. The other supplies or commissions task examples that already cover the desired behavior. The publisher should decide whether it wants to grant use of its bank alone or also commit staff to the new work. The buyer should decide whether the distinctive source is essential, then inspect a representative completed unit before comparing quantities.

Bottom line: Innodata is the stronger starting point when the missing work is expert task creation or answer checking. Its benchmark offer also fits a need for evaluation design. Rightmark fits when access to particular proprietary material is the missing input and its existing structure, together with assigned preparation work, can meet the task.

Ask both: "Show one representative unit with its input, expected answer and scoring rule. Which parts arrive with it, and which must the buyer or owner still create and check?" A file count without that distinction leaves the purchase undefined.

2. What does the buyer pay, and what does the owner receive?

Two offers can quote the same number of questions and cover different amounts of work. One may license the existing bank; another may include checking answers, rewriting prompts and preparing scoring guidance. Owners also need to separate payment for the asset from payment for their additional time.

How do Rightmark and Innodata compare for a defined model task? (13)
QuestionRightmarkInnodata
What is the buyer's price basis?On standard licenses, the buyer pays Rightmark's fee on top of the owner's payout. The total and fee amount need a proposal.Managed-work prices and billing units: not published. Request the unit and inclusions.
What does the owner keep?Standard licenses have no deduction from owner payout for Rightmark's fee. Premium/bespoke wording says owners keep 75% to 85%, agreed in writing.Its privacy policy acknowledges supplier and licensing inquiries; it publishes no standard owner payment schedule.
Who pays for preparation?For company archives, extraction, scrubbing, transfer and fees are agreed in writing.Separate the existing-data charge from requested authoring, checking and transfer work in the quote.
What is the commitment?The license defines its term. A universal minimum commitment is not published; payment-trigger terms are developing.Minimum commitment and payment milestones: not published.
What if an item is rejected?Standard remedy terms are developing. Agree correction, replacement, refund and extra-work treatment for the deal.The evaluation offer page does not publish correction, replacement or refund terms. Resolve them for the engagement.

Rightmark's public fee statement is specific: "We take nothing from your payout on standard licenses: the buyer pays our fee on top. On premium and bespoke deals you keep 75% to 85%, agreed in writing before any sale." The standard rule describes who pays the intermediary fee. It does not say the owner receives all the buyer's total spend, or that every extraction and preparation expense is included.

For a bespoke proposal, the percentage also needs a calculation base. Ask whether it applies before or after separately charged work, and which deductions sit outside it. Agree the precise calculation base and treatment of additional work in writing. Without those fields, an owner cannot compare a percentage with a fixed payout meaningfully.

Use this blank aid for the same proposed scope. Keep a charge on one line only: if preparation is already included in the license amount, mark it included instead of adding it again.

How do Rightmark and Innodata compare for a defined model task? (17)
Proposal fieldProposal AProposal B
Counted delivery unit and quantity____________
License amount and included work____________
Buyer-paid intermediary fee, if separate____________
Separately billed preparation and transfer____________
Owner-funded charges or deductions____________
Acceptance rule and payment trigger____________
Total buyer bill____________
Owner proceeds after agreed deductions____________

For the invented question bank, keep the existing-asset license on its own line. Add new answer checking or rubric work only where the proposal requires it, whether the publisher, buyer or another preparer performs that work. Calculate a per-unit figure only after both proposals count the same accepted unit and include the same work; otherwise the arithmetic makes unlike purchases look comparable.

Bottom line: Rightmark gives owners a published intermediary-fee basis to examine. It establishes neither a higher owner payout nor a lower buyer bill. Choose on the total cost of getting the agreed deliverable into use, and keep the owner's proceeds visible alongside that total.

Ask both: "For this exact deliverable, what is the total buyer bill, what reaches the owner, which preparation charges can change, and what happens to payment if an item is rejected?" Ask for the calculation, not just a percentage.

3. Can the owner grant the use the buyer actually needs?

A useful answer can sit inside a collection whose rights differ item by item. The buyer needs permission for its intended use, and the owner needs a contribution narrow enough to support. Finding a promising source is the start of that work.

How do Rightmark and Innodata compare for a defined model task? (23)
QuestionRightmarkInnodata
Who grants the rights?The buyer signs with the owner, or on code routes with a licensing partner selected per deal.Identify the licensor of the source material and any newly created components.
Which uses are covered?Document training, retrieval and display rights are scoped and priced separately; the license states whether derived datasets are included.The fine-tuning offer page does not publish a dataset license. Obtain the grant for the actual deliverable.
What must the owner exclude?The document route limits licensing to owned rights; freelance work, wire copy and agency photos may need exclusion.List third-party elements in any owner contribution and establish rights in commissioned additions.
What does sample access permit?The public buyer NDA limits portal material to deciding whether to license. Agree the specific model test in writing.The website privacy policy is not a dataset license. Obtain sample-processing and retention permission separately.
What supports review?A route-specific ownership pack built from owner documents and reports, shared with counsel under NDA.Request provenance and permission evidence for the source and for any created answers or rubrics.

An anonymized Rightmark pre-deal exchange shows why scope matters. An expert-conversation archive was narrowed to recordings whose guest releases permitted third-party licensing; the rest stayed out of consideration pending updated releases. That was a correction to the proposed contribution, before any evidenced delivery or payment. It is a useful discipline for an owner: identify the material that can support the requested grant, then describe that smaller asset accurately.

For the question-bank owner, the equivalent exercise is to separate original questions, third-party illustrations and answers licensed from contributors. A buyer asking to train on the questions may also want to display explanations in a product. Put those uses into the request individually, so the owner can establish which it can authorize and which require an exclusion or further permission. The data card template provides a place to record scope and known gaps; it does not resolve them by itself.

Bottom line: Rightmark's documented owner-and-asset route fits a purchase where establishing rights in particular existing material is central. Innodata's expert-created data remains relevant when those are the outputs needed. In either case, the agreement must cover each component the buyer intends to use, including additions made during preparation.

Ask both: "Who licenses each source and created component, which uses and derived datasets are covered, and what must the owner exclude or obtain permission for?" A privacy notice does not answer a question about a dataset grant.

4. Who approves the scope, and what stays out of training?

Signing can authorize more than one immediate transfer. Owners need to know what permission continues afterward; buyers need to know which sample or evaluation material can enter which system. Those are separate controls, even when both appear in the same proposal.

How do Rightmark and Innodata compare for a defined model task? (30)
QuestionRightmarkInnodata
What does signing authorize?On code routes, signing covers future non-exclusive sales.Identify which uses or further transfers need fresh approval in the requested agreement.
What can an owner withdraw?Code not yet licensed can be withdrawn; that statement does not retract existing licenses.Owner withdrawal terms are not published in the inquiry policy. Request the asset-specific terms.
What scope does the buyer approve?Families, filters, formats, de-identification, volume, exclusivity and delivery cadence are put in writing.Specify the selected deliverable, requested changes and sign-off points in the proposal.
Can a sample enter a model run?Agree the exact test first. The NDA also prohibits attempts to identify people in samples.Request explicit processing, access, retention and deletion boundaries for the test.
How is an evaluation kept separate?Request evidence of separation from training and from the tested system's access; private-source status alone does not establish it.Request the same evidence for a prepared pack or commissioned evaluation.

NIST CAISI distinguishes training contamination from answer exposure during an evaluation. Its examples show that a tested system can encounter solution material at test time even when the task was absent from training. This makes the access question concrete. If the task asks for a diagnosis, record where the reference answer lives, who may inspect it and what the tested system can retrieve. A clean split of files is incomplete if the system can still reach the answer through another permitted tool or document.

For an owner, write down the authorization that survives signing before deciding whether repeat licensing is attractive. Rightmark's published code route can suit an owner comfortable authorizing future non-exclusive transactions within the signed scope. Someone who wants a fresh veto before every transaction needs that different arrangement agreed explicitly. The archive or document route needs its own terms; the code rule does not answer for it.

Bottom line: Rightmark makes a concrete code-authorization tradeoff public. For the buyer's evaluation controls, choose the proposal that documents both the permitted test and the required separation. Keep commercial permission, preparation access and test-system access as distinct decisions.

Ask both: "What may you do after I sign, what can I still withdraw or approve, and how are evaluation questions and answers kept out of training and away from the tested system?" Ask for the actual control and its responsible owner.

5. Does the offer cover your domain, format and task?

A label such as engineering or coding narrows the search, but it does not establish the needed behavior. A buyer might need troubleshooting rather than explanation, or changes across a codebase rather than isolated functions. Owners should describe the structure already present before agreeing to expand it.

How do Rightmark and Innodata compare for a defined model task? (37)
QuestionRightmarkInnodata
Where is the distinctive content?Source particular code, archives, documents or question banks to the buyer's written specification.Coding datasets cover generation, debugging, software reasoning and benchmark-style evaluation.
What structure accompanies documents?Worked solutions, rubrics and graded answers can stay with questions; preserve the structure useful to the task.Question generation and answer verification are offered. Check which is commissioned.
Which formats are supported?Specify the asset's actual format; the documents page describes TMX translation-memory exports with language and domain metadata.Fine-tuning data spans text, image, video, audio and code, across languages.
What needs adaptation?Distinguish filtering existing material from new authoring; agree added non-code work separately.The catalog offers adaptations, extensions and domain expansion.

Take the format question down to one record. In the invented bank, the buyer should be able to see how a question is linked to its answer and whether an explanation belongs to the same revision. If diagrams carry essential information, an export containing text alone changes the task. If a source uses a different vocabulary from the intended users, assign the mapping work instead of silently treating the domain label as sufficient.

For code, specify whether the desired unit is a repository, a change with its context, or a prepared question. These are different requests even when all three concern debugging. A codebase's owner-supplied CI report can inform inspection, but it does not replace the acceptance check for a newly constructed task. Rightmark's described in-preparation code lots remain unavailable; a task-shaped request still needs sourcing confirmation.

Bottom line: Innodata is a strong fit when the required work is expert preparation across languages or formats. Rightmark fits when distinctive proprietary material is essential and the buyer can identify the useful structure within it. The owner's corresponding choice is how much new work, beyond licensing the existing asset, to accept.

Ask both: "Can you show a sample matching this domain, language, format and task, and identify what is already available versus what must be sourced, created or adapted?" Keep each unresolved requirement visible in the scope.

6. Who turns the material into an accepted delivery?

The handoff can fail even when everyone agrees the source is valuable. Someone must extract the material, preserve its structure, perform the agreed checks and decide whether the delivered unit passes. Owners should see those duties before committing their staff; buyers should know what remains on their own team's calendar.

How do Rightmark and Innodata compare for a defined model task? (44)
QuestionRightmarkInnodata
Who prepares the handover?Written scope defines filters, formats and cadence. Assign the inputs required from owner and buyer.Specify the dataset handover separately from any commissioned service output.
Who extracts and scrubs?Company-archive responsibilities and charges are agreed in writing. Agree the non-code work split for the proposed delivery.Identify any access to private systems, extraction work and masking in the service scope.
Who creates and checks answers?Existing solutions can accompany a question bank. Agree additional authoring and checks separately.Expert-created fine-tuning data includes offered question-generation and answer-verification tasks.
Who runs the evaluation?A graded batch does not itself specify a managed model evaluation.Red-teaming services include expert prompt writing and model-response rating.
Who accepts and corrects it?Name the checker and task test. Standard remedy terms are developing.The evaluation page publishes no standard acceptance or correction commitment. Agree it for the chosen scope.
When can work start?Confirm sourceability, volume and timing; sample timing belongs in the written scope.Confirm the selected deliverable and timetable before scheduling dependent work.

Rightmark's historical archive correspondence shows part of this handoff. The team asked owners' system admins for approximate counts and first-record years, separated company-owned records from customer content, and held the latter while checking export authority. In another exchange, it prepared buyer-specific one-pagers with sources and dates for the figures, then requested admin screenshots to substantiate volumes. These were pre-deal scoping and evidence-gathering steps; the records do not show a transformed archive or accepted delivery.

That division is useful to make explicit in a proposal. An admin who can supply a count may not be the person who can export the records. A subject expert who wrote an answer may not be available to check its conversion into a model task. Name both roles where needed, and specify the output of each handoff: a manifest, a scrubbed sample, a corrected answer set or a scored report.

Acceptance should also distinguish three questions. Did the promised files arrive? Do the answers and scoring rules meet the agreed checks? Does the buyer's model perform as hoped? The first can be true while the second remains disputed, and neither establishes the third. Set the acceptance unit and the remedy around the work actually purchased, with responsibility for any additional iteration agreed before it begins.

Bottom line: Innodata fits a buyer commissioning expert prompt creation and response assessment. Rightmark fits when sourcing and licensing the asset is the missing work, with the remaining preparation assigned and costed. A preparation task without an owner, price or acceptance rule is still an open part of the purchase.

Ask both: "Who handles extraction, rights review, answer creation, validation, delivery and corrections, and which tasks are included in this price?" If the owner must provide new work, put those hours or deliverables beside the license terms.

7. What evidence supports the offer you are considering?

A sample can show what a record looks like without proving that a delivery was accepted. Owners need an equally specific distinction between advertised fee terms and what a comparable license actually paid. Request evidence for the work and outcome that matter to this proposal.

How do Rightmark and Innodata compare for a defined model task? (52)
QuestionRightmarkInnodata
What does the public artifact show?Ownership-pack examples are illustrative, explicitly not from a real dataset.Published benchmarks describe scenarios and scoring methods. Inspect the relevant method.
Is there an operating example?An anonymized code record documents a qualification correction and progression to buyer review.An anonymous vendor case describes model-output assessment.
What proves fit for this buyer?A permitted representative test of the proposed material against the intended task.The same task-specific test or assessment of the commissioned method.
What proves owner proceeds?A comparable payment record would be needed.A standard owner payout schedule is not published in the inquiry policy. Request terms and appropriate evidence.

The qualification record shows a concrete check. In an anonymized internal code review, a corrected ownership document had dropped an earlier never-public warranty. Staff requested its restoration or disclosure of prior publication. Later notes recorded restored signed wording and qualification for buyer review. The evidence here is the notes and stage history, not an inspection of the signed documents or a completed-sale record. It shows how a changed statement was caught before the next review stage.

Innodata's anonymous evaluation case describes a different job: assessing prompts and corresponding model-generated outputs, with the provider and client refining the rating guidance together. The vendor-authored account concerns evaluation work; it supplies no measured catalog-dataset result. For a buyer commissioning assessment, the useful next question is how raters handle a disagreement. Ask to see the criteria, an example of a disputed judgment and how the final decision is recorded, with confidential material removed where needed.

Bottom line: Innodata provides the more relevant published workflow example for commissioned model-output assessment. Rightmark's available records support qualification and pre-deal scoping, while its completion count lacks task-level detail. Favor evidence that matches the proposed work; request a different record when the claim concerns acceptance, model performance or payment.

Ask both: "Show evidence from a comparable engagement: what was delivered, how was it accepted, and which record supports the claimed result or owner payment?" A process description can explain the approach, but it cannot supply an absent result.

Innodata alternatives: what Rightmark does differently

  • Source particular private material. Rightmark's buyer process starts with a specification and checks title and consent while sourcing from owners. Families, filters, formats, volume and delivery cadence go into the written scope, so a buyer can describe the required source precisely before availability is confirmed.
  • Retain the answers already in the asset. The document route keeps worked solutions, rubrics and graded answers with each question. That lets the buyer inspect existing structure before deciding which additional authoring or checks to commission.
  • Assemble ownership evidence for the route. The ownership pack draws on owner documents and reports and is shared with counsel under NDA. Its contents vary with the asset; the pack is a basis for review rather than legal clearance or a substitute for the license.
  • Separate sample access from the eventual grant. Buyer samples are for evaluation only under NDA, with the proposed test separately scoped. The eventual license identifies the covered material and derived-data treatment, while document training, retrieval and display rights are considered separately.
  • State the intermediary-fee basis. On standard licenses, Rightmark's fee is added to the buyer's bill rather than deducted from the owner's payout. Premium and bespoke deals have the separate published 75% to 85% owner-share statement, agreed in writing, with the calculation base and additional charges checked for the proposal.

When is Innodata the better choice?

Choose Innodata when you need expert-created task data, or a designed evaluation with scenario-specific scoring, and the proposed work matches the behavior your team needs. The essential purchase may be the prepared deliverable rather than access to one particular source archive.

An offered pack can also be extended or adapted. If its sample already covers most of the task, specify the remaining changes and inspect the resulting unit before committing to the full scope. Keep the definition of done stable while comparing proposals: a polished example and a reproducible acceptance check answer different questions.

An owner evaluating a concrete contribution proposal should compare the asset rights with the new labor requested. Licensing an existing bank, writing fresh questions and adjudicating disputed answers are different contributions, even if one agreement covers them. Innodata's acknowledgment of supplier and licensing inquiries leaves room for discussion; it does not establish an open paid-contributor program or standard asset-owner terms.

When is Rightmark the better choice?

Choose Rightmark when the required content belongs to a particular proprietary source and the parties can define its rights, delivery scope and remaining preparation. A company archive's context, a document collection's depth or an existing question bank's worked answers may be the reason for the purchase. Sourcing that asset and documenting permission are then central parts of the job.

This route is especially relevant when useful task structure already exists and the buyer has someone responsible for assessing and completing it. A publisher should identify the material and rights it can contribute, then price any requested new work separately. Rightmark's fee statement gives that conversation a starting point, while the proposal must settle actual proceeds, preparation charges and the payment trigger.

Code owners should also be comfortable with the published authorization model: future non-exclusive licensing within the signed scope, with withdrawal limited to code not yet licensed. For buyers, any bundled expert preparation, model evaluation or correction commitment still needs an explicit scope. Agree those preparation and checking details in writing; standard remedy terms are developing.

Consider a third option if...

You already control the material and the necessary use rights, and an internal team can do the remaining work. Compare that cost with purchasing another source or a managed engagement. An owner with a known buyer may likewise be able to arrange the license directly, provided the parties can handle the required documentation and preparation.

You need a particular source and specialist preparation from separate providers. Put the preparer's access, allowed transformations, rights in created components and acceptance duties into the respective scopes.

The observations you need have never been recorded. In that case the next brief should describe collection: which events, people or conditions must be captured. If the owner cannot support the requested rights, resolve the gap or exclude the material before choosing any route.

Relevant reading: Off-the-shelf vs custom training data examines the separate decision between existing observations and new collection.

The decision is still yours

Choose the work you need supplied: a particular source asset with permission to use it, prepared examples for a defined task, or expert work to create and evaluate them. For an owner, the same choice determines whether the commitment ends with licensing existing material or includes additional contribution work.

Take the Ask both questions into the proposal discussion. Put the deliverable, permitted use, preparation responsibilities, acceptance rule and total charges in the same written request, with deductions and the payment trigger visible to the owner. The useful offer is the one that resolves those questions for your actual material and task.

Have data to license?

Answer a few questions about your data's volume, history and rights for a free indicative range and likely licensing routes. The appraisal needs no files or system access, and the estimate is not an offer.

Appraise your data

Sources

What brings you to Rightmark?

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

What brings you to Rightmark?