Best licensed text dataset providers for LLM training in 2026

Does the corpus contain the language and knowledge your model needs?

Describe the useful slice: subject, language and dialect, document type, period and context retained per record. A chapter preserves an argument; a translation pair preserves the relationship between two versions; a worked answer preserves the steps behind a result. An owner should identify which works satisfy that request and which stay outside it.

Ask: “Show the collection's breakdown by language, subject and document type, then give us a sample from the part our model needs most.” Check that the sample comes from the quoted collection. A total corpus size offered in place of the specialist subset is a reason to keep looking.

Which training rights come with these files?

Have the contract reviewer match the proposed grant to the actual works, entities and model uses. Owners should check author, freelance, client and user rights, alongside duration, exclusivity and what they can license elsewhere. Keep retrieval and display out of the request unless you need them too.

Ask: “Show us where this agreement permits our stated training use, what we may do with the sample, and which works or uses are excluded.” A general assurance that text is licensed leaves those questions open. Agree permission for a sample test before running it.

How much usable text is included in the quote?

Request counts before and after the agreed filtering, redaction and duplicate removal. If the quote uses tokens, specify the tokenizer and counting method. Inspect a few records before and after preparation: did a news article lose its date, or a question lose its solution? Owners need the same clarity when exclusions change the payable volume.

Ask: “Show the counts, preparation rules and example records behind this price.” State whether duplicate checks cover only the collection or an authorized reference set too. Raw-volume pricing is workable when the basis and expected delivery are explicit; an unexplained difference between billed and usable volume is the problem.

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

Compare the same corpus, rights, preparation and delivery. List license charges, mandatory services and recurring costs separately. For an owner, record the payment amount or share, its calculation base, deductions, payment trigger and retained rights. A larger percentage alone cannot tell you which offer leaves more money with the owner.

Ask: “For this exact scope, what is included in the total, what can add a charge, and when is payment due?” Owners should add: “What amount reaches me, what is it calculated from, and what makes it payable?” An undefined fee base or unpriced essential preparation prevents a useful comparison.

Can the provider deliver the collection you reviewed?

For a fixed training date, I would favor a matching collection available to inspect now over material that still needs sourcing. New collection earns its place when it fills a specific gap and the remaining work fits the schedule. Establish which version, volume and fields will arrive, who reviews the full batch and how mismatches are resolved.

Ask: “Which material exists now, which work remains, and how will delivery be checked against this sample and specification?” An honest collection plan is useful. A promise of ready data with no identified collection, or no answer about rejected records, is not enough to schedule a run.

“Not published” means the reviewed public offer does not state that term; request it in the proposal. An owner share is not a buyer price. Compare the same corpus, rights and included work before comparing totals.

Rightmark: best for sourcing private text archives

At Rightmark, we source and license privately held text against a buyer's requirements. Our documents offer covers backlists, news archives, translation memories, courseware and forum threads. Through the buyer process, we put the families, filters, formats, volume, exclusivity and delivery cadence in writing. Start with the material the model needs: a translation memory should retain its language pairs; a question bank should keep the worked solution attached. Owners begin with the archive and the rights they can grant. For news, keep dates and bylines in the scope so reviewers can compare the promised coverage with delivered records and identify excluded material.

Our documents page treats training, retrieval and display as “Separate rights, priced separately.” The buyer terms say: “Nothing is licensed until the owner and your company sign a license agreement.” The homepage states: “We take nothing from your payout on standard licenses; the buyer pays our fee on top.” Owners pay nothing upfront; premium and bespoke deals leave owners 75% to 85%, agreed in writing. Ask for the share's calculation base and any deductions, and compare the buyer's complete bill separately. The proposed payment should make sense for the exact text and rights being licensed.

Allow time for archive qualification and rights work. If a suitable catalog collection is already inspectable and your training date is fixed, compare it first. Our sample stage provides representative scrubbed material under NDA for the licensing decision; agree any model test before running it. The signed license governs the deal. Put acceptance, correction work, extra charges and the owner's payment trigger there.

Choose Rightmark to source a specific private archive with written rights and fees; compare an inspectable catalog first when a matching collection and fixed start date matter most.

Created by Humans: best for author-approved book licensing

Created by Humans offers book licensing across rights holders through one agreement and invoice, with training, reference and transformative rights available. For a training purchase, turn the reading list into a delivery specification: titles and editions, language, subject coverage, complete works or selected portions, and identifying metadata. A familiar author can make a collection worth investigating, but the sample still needs to contain the kind of writing the model needs. Check whether a proposed subject collection contains sustained explanations, reference material or storytelling prose before treating those books as equivalent learning material.

Authors can approve negotiated agreements and track payments. Walter Isaacson's library includes training/reference titles and an unlicensed work. That makes the work itself the sensible unit for checking permission: selecting an author is only the beginning. Isaacson, an investor, advisor and founding author, describes the aim as “control over how our work is used by AI companies” on the homepage. Owners should use the approval step to examine the requested uses, compensation basis and deductions together. Buyers should match the final list of works to the actual training grant, then reconcile that list with the delivered files.

Choose this route for books. A requirement centered on support exchanges, translation memories or newly written task answers calls for a different source. Before committing, request a representative extract, file specification, delivery date and permission for the proposed sample tests. Have the offer state the buyer's price, the owner's proceeds and the treatment of incomplete or unusable files. If an owner plans to opt out later, ask what that changes for future offers and for licenses already granted. A platform setting and the term of a signed agreement need to be understood together.

Choose Created by Humans for a book corpus whose titles and training rights match; use an archive or collection specialist when the required text is not books.

AP Intelligence: best for first-party news archives

AP Intelligence supplies AP-owned current and archival reporting, including multilingual text. Its Source Data offer includes model training. This is a concrete route for a team that needs reporting on events. Begin with the period, subjects and languages the model must learn from, then request a coverage breakdown and an extract from that slice. A corpus dominated by one period or topic could leave a different part of the requirement unfilled. Decide whether you need a bounded historical purchase or an ongoing supply before comparing the commercial proposals. AP's announcement describes the source material.

AP advertises metadata, tagging, taxonomy and feed or API delivery. Ask how the package identifies a story and its updates: counting successive versions as separate documents may be appropriate, but that choice should be deliberate. For its exploration of generative AI in news, OpenAI licensed part of AP's text archive in July 2023. In the announcement, then-COO Brad Lightcap said the access and feedback “will help to improve the capabilities and usefulness of OpenAI’s systems.” For your purchase, inspect the schema and text completeness alongside the proposed training grant; decide how corrections and later versions should enter the delivered collection.

AP is the news-source choice here. Books, support conversations and expert-written answers require another corpus. An independent publisher seeking someone to license its own archive has a separate representation question to ask. Request a proposal naming the selected news slice, model uses, usage limits, delivery cadence and complete fee, with acceptance and remedies attached to that scope. Price a continuing feed separately from a historical batch so recurring costs remain visible. AP's website terms cover site use; obtain the AI-data license for the purchase itself.

Choose AP Intelligence for structured AP news text; use another source for books or task examples, and a separate representation route for an unrelated owner's archive.

Defined.ai: best for catalog and custom text options

Defined.ai combines catalog text with custom collection. Its text page lists English books and Japanese legal templates, plus conversational and labeled data. A buyer can therefore begin with a specific choice: inspect an existing collection or define the missing material to commission. Keep the learning task visible during that decision. Long passages, short classification examples and multi-turn exchanges ask different things of a model. Request a sample in the intended format, with enough surrounding context to judge whether the text addresses the task, before commissioning more volume.

The text page describes document, passage, instruction and conversation schemas, with domain, locale, source-type and rights metadata. Ask which fields and validation steps are included in the selected package. Its DLA permits commercial trained-model use subject to its terms, but restricts data commercialization and recognizable contributor outputs. In a testimonial about video, audio and word datasets, Saurabh Saxena, Head of Technology, VP R&D (EmotionAI), says: “We are delighted by their rigor and reliability.” For the buyer, the useful next step is to inspect the text sample and agree how errors or missing fields will be handled in the full delivery.

The partner program offers marketplace licensing after quality and rights review, with commercial terms approved by the partner. Owners should request the proceeds calculation, reporting and payment trigger for their collection, separately from the buyer's quote. Buyers planning to deliver source text to clients need prior written authorization; the DLA supplies data as is. If redistribution is essential, resolve it before evaluating the price. Also confirm whether the proposed collection already exists or needs new preparation, and put its version, delivered volume and acceptance checks in the order.

Choose Defined.ai to compare catalog text with a custom brief; choose another route if required source-text redistribution cannot be authorized.

Shaip: best for Arabic instruction data and specialist text

Shaip, part of Ubiquity, offers English and Arabic conversational Q&A and instruction text, including Modern Standard Arabic, Saudi, Emirati, Egyptian and Levantine dialects. Its catalog also covers news and academic publisher text. For a regional-language assistant, begin with examples from the intended dialect and task. For scientific learning, begin with the relevant subject and full text. Those are separate selection problems even if the buyer can obtain both from one supplier. Ask a reviewer familiar with the target language or subject to inspect the proposed slice before choosing its volume.

The catalog describes news bylines, dates, headlines, bodies and topic tags, academic API delivery, and samples before licensing. These give the buyer specific fields and a delivery route to test. In its October 27 announcement, Shaip distinguishes existing catalog material from collection and annotation used to extend it. Have the quote separate that additional work so a request for one missing dialect does not become an undefined preparation project. An unnamed Senior Applied Scientist at Oracle praises the “support and professionalism your team has consistently provided” in a general service testimonial. Judge the selected corpus through its own sample and agreed checks.

The announced catalog licenses are non-exclusive. If keeping the same material away from competitors matters, seek a separately negotiated exclusive grant or another route. For a custom requirement, describe the behavior and examples missing from the catalog, then agree what must be created and when. Owners offering an existing archive should establish whether the proposal licenses their text, represents their rights or commissions new work; ask for compensation and downstream uses on that basis. Customer refunds follow the contract; otherwise payments are final and non-refundable under Shaip's terms. Set acceptance criteria and any remedy before approving the order.

Choose Shaip when its dialect and subject collections fit a non-exclusive purchase; seek a separate grant or another route if exclusivity is essential.

Innodata: best for expert-written reasoning and fine-tuning tasks

Innodata offers prebuilt training and evaluation tasks. Its Math QSA pack contains English graduate mathematics in Excel, separating question, worked solution and final answer across continuous analysis, discrete structures and optimization. That structure gives a reasoning-data buyer something precise to inspect. Read the solution as carefully as the final answer: decide whether its steps demonstrate the skill the model should learn and whether the difficulty matches the intended users. Ask how the record fields should enter your training pipeline, how many usable records are offered and which sample tests are permitted.

The Math QSA page describes expert authoring and SME complexity validation before benchmarking. Its fine-tuning service covers expert-created dialogue and full-length content. The collection service includes prompt-response and expert Q&A work with schema and delivery designed to the requirement. Separate available records from new authoring or extra review when requesting a quote. Sparrow Computing needed annotations for an inspection model; founder and ML engineer Ben Cook says, “Innodata really adds value to our operations” in its testimonial. For a reasoning corpus, use the pilot to check correctness and difficulty, then resolve failures before expanding the order.

The inspected math pack is a narrow starting point. If you need another language, elementary instruction or a publisher's original prose, request a matching collection or commission the missing scope. Do not approve a pack merely because its file structure fits the pipeline. Specify the subject mix, accepted-record count, delivery date and treatment of flawed examples. Owners should distinguish licensing an existing archive from paid authoring and preparation services: compensation for writing new questions is a different bargain from granting use of an existing backlist. Ask who may use the source text and the resulting task examples, and which rights and payment remain with the owner.

Choose Innodata for structured reasoning tasks whose subject and level fit; seek another collection or commission the scope when you need different languages or original publisher prose.

LXT: best for commissioned multilingual and domain-specific text

LXT, whose group includes clickworker, collects or creates conversational, localized and specialist text. Its LLM collection service covers technical manuals, scientific publications and regulatory documents. This route suits a buyer whose brief needs more than a language label. State the professional setting, regional expressions and tasks the material must represent, then ask which records will be newly authored and which will come from existing sources. For either kind, identify the training permission being offered. An owner bringing an archive should establish whether the engagement is licensing that archive or paying for new contribution work.

The text workflow sets contributor guidelines, matches domain and language skills, and runs a pilot before production, with gold tasks, expert review and automated checks. Inspect the pilot's errors and coverage before increasing volume. Dubber needed multilingual transcription benchmarks before a launch; LXT collected and transcribed speech and supplied samples by language in its published case. Iain McCowan, Director AI at Dubber, says, “LXT is a key partner in our product expansion efforts” on the language page. LXT also promises to work with dissatisfied customers toward expected quality. Agree the actual acceptance checks and remedies in the proposal.

Text pricing depends on volume, domain, languages, expertise, creation versus collection and turnaround; minimum volumes may apply. Request the complete scope for preparation, delivery and refreshes, then record how rejected examples affect the bill. Owners need their own compensation and reuse terms, whether contributing new writing or licensing existing material. Commissioning text leaves collection work on the schedule. If the training run requires a particular archive immediately, choose an inspectable existing collection unless LXT confirms that exact ready supply. For a commissioned project, make pilot approval the point at which the parties confirm the remaining work, final delivery and cost.

Choose LXT for text commissioned to a language or domain brief; use an existing licensed corpus when an immediate, fixed-archive requirement leaves no time for collection.

Shortlist two or three providers for the same text requirement

Write one specification covering language, subject, source type, period, usable unit, format and training use. Narrow to two or three providers that can meet it with available material or an explicit collection plan. Record the work still needed beside each name. Owners should also choose the arrangement they want: archive licensing, buyer representation, preparation services or paid authoring.

Send the same buyer request brief to each candidate. Keep a separate comparison sheet with five rows: corpus fit, training rights, usable delivery, buyer cost and owner proceeds, and readiness. For each row, record the answer, its supporting document or sample, the unresolved point and a pass, clarify or fail decision. A provider that does not offer the owner's proposed arrangement can be ruled out on fit without judging its buyer service.

Ask every shortlisted provider the same seven questions

1. “Which exact text would I receive, and what is outside the package?” Match the languages, subjects, periods, record unit, metadata and exclusions to the sample and quote. A red-flag answer is: “The package is all English text; we choose the actual sources after you commit.”
2. “Who can authorize this training use, and which material does the written permission cover?” Establish the rights chain, model use, duration and restrictions. A red-flag answer is: “The text is public online, so no separate permission needs checking.”
3. “Which tests may I run on the sample, and who approves access and retention?” Name inspection, training, fine-tuning, benchmarking, embeddings or derived datasets separately, with people, systems and an end date. A red-flag answer is: “The NDA lets you run any model test and keep the results indefinitely.”
4. “What would make the delivery fail acceptance, and what happens to rejected material?” Agree coverage, correctness, duplicates, metadata, reviewers and any correction, replacement, fee or refund. A red-flag answer is: “We count everything we send as accepted; errors and missing fields are your problem.”
5. “What is the complete price for this scope, and what could add to it?” Include the billable unit, minimum, rights, pilot, preparation, delivery and refresh. A red-flag answer is: “This is only the license price; we tell you the required preparation charges after delivery.”
6. “What exists now, what must be collected, and what happens if the pilot stops?” Establish milestones, delivery, remaining charges and treatment of files, access and permitted derived material. A red-flag answer is: “Pay for the full order first; delivery and exit conditions will be decided later.”
7. “If I supply text, which rights do I keep, what amount is due to me, and when is it paid?” Distinguish archive licensing, new writing and preparation, then pin down compensation, deductions, downstream uses and approvals. A red-flag answer is: “Agree to the uses now; we will decide your share and deductions after the buyer pays.”

These red-flag answers are illustrative, not quotations from the listed providers. Record a concrete unresolved item when a real answer is incomplete.

Agree a 30 to 60 day pilot with written limits

Propose a 30 to 60 day pilot before the larger commitment. Set the start event and end date: for example, the clock could begin when the scoped sample arrives with its required permissions. Agree whether the test uses a limited existing corpus, a paid sample or a small collection assignment, and specify the price and any minimum.

Use the sample-pack checklist to record the review question, material boundary, owner approval and transfer. Add the exact volume, fields, exclusions, permitted tests and reviewers. For task examples, check the questions and solutions; for source prose, check subject coverage, period and completeness. Write the acceptance thresholds or review procedure before testing. Include a model-performance experiment only when the parties agree its method and permitted uses.

Define underperformance and the response: who reviews a disputed finding, whether work is corrected or replaced, the time allowed, extra charges and any refund. Record when access ends, what must be returned or deleted, and how permitted derived artifacts are handled. Owners should approve pilot uses and compensation before releasing material.

If a provider resists, ask which minimum license, setup cost or confidentiality constraint prevents the pilot. Compare that constraint with the purchase. An alternative evaluation can be reasonable; refusing to define any evaluation, acceptance or exit process leaves you without a controlled decision. At the end, choose whether to proceed, revise the scope or stop against the written criteria.

The proposal worth advancing connects suitable text, explicit training uses, a representative sample and delivery that fits the project. Compare the complete price of that material and work. An archive, prepared task examples and commissioned writing can each win for a different requirement.

For owners, the decision is whether the allowed uses, retained rights, compensation and payment trigger justify the grant. Request a revision when a material point remains unresolved, before agreeing to the use or purchase.

Put the text, intended use and review requirements into one brief for your shortlisted providers.

Use the buyer request brief

Source dates distinguish publication, publisher updates and archive captures. The text and partner pages were rechecked November 3, 2026.

• Rightmark documents. Updated 2026-10-10.
• Rightmark buyers. Updated 2026-10-21.
• Rightmark fees. Updated 2026-10-26.
• Rightmark terms. Updated 2026-10-28.
• Request brief. Updated 2026-10-18.
• Sample checklist. Updated 2026-10-18.
• Created by Humans books. Archived 2026-07-31.
• Isaacson library. Archived 2026-05-10.
• Created by Humans introduction. Archived 2026-04-22.
• AP Source Data. Published 2026-10-19.
• AP Intelligence launch. Published 2025-11-06.
• AP and OpenAI. Published 2023-07-13.
• AP terms. Published 2024-02-09.
• Defined.ai text. Archived 2026-06-12.
• Defined.ai partners. Archived 2026-10-24.
• Defined.ai license. Effective 2022-11-30.
• Defined.ai testimonial. Archived 2026-10-15.
• Shaip text. Published 2026-09-25.
• Shaip catalog announcement. Published 2026-10-27.
• Shaip testimonial. Published 2024-06-18.
• About Shaip. Published 2025-01-13.
• Shaip terms. Published 2020-12-07.
• Innodata catalog. Published 2026-04-22.
• Innodata Math QSA. Published 2026-09-12.
• Innodata fine-tuning. Published 2025-01-24.
• Innodata collection. Published 2026-06-26.
• Innodata testimonial. Published 2023-09-08.
• LXT text collection. Published 2025-08-25.
• LXT LLM collection. Published 2025-08-29.
• LXT language testimonial. Published 2022-12-20.
• About LXT. Published 2022-06-23.
• LXT Dubber case. Published 2022-06-27.
• LXT quality support. Published 2023-01-13.