Contract and clause extraction datasets A buyer guide to contract corpora, clause labels, review rationale, public benchmarks, permissions, commercial scope and defensible evaluation. Contract datasets connect agreements with the provisions, relationships and review judgments a legal or business model must understand. They may contain original text, clause spans, document metadata, question-answer pairs or reviewer explanations. A useful source preserves agreement families, amendments, jurisdictional context and uncertainty rather than treating every paragraph as an independent example. CUAD provides a public clause-review reference, while private contract sourcing needs its own authority, confidentiality and intended-use review. Compare the legal meaning of labels and the evidence behind them, not only contract counts. This guide concerns dataset design and sourcing; it does not provide legal advice. Dataset purpose Contract and clause extraction data Availability Source and commercial scope reviewed for each project Reviewed 12 November 2026 Choose the contract task before the corpus Contract classification, clause extraction, obligation tracking, risk review and question answering have different data requirements. A clause detector may need spans and a taxonomy, while an obligation system needs the responsible party, action, condition, deadline and exceptions. A retrieval model may need complete documents and evidence references. Define the output that will count as correct before deciding whether a large text archive is suitable. A commercial contract is also part of a relationship. Schedules, order forms, amendments and referenced policies can change how a provision operates. A paragraph taken out of that context may be impossible to interpret reliably. Preserve document family and version identifiers, and state whether the source includes the material needed to resolve a reference. A corpus of isolated clauses can still be useful, but should be described as such rather than presented as complete agreement understanding. What CUAD contributes to the comparison [The Atticus Project’s CUAD page](https://www.atticusprojectai.org/cuad/) lists 510 contracts, more than 13,000 labels and 41 clause types under CC BY 4.0. It provides a concrete public reference for clause-focused annotation. Those counts do not make it a representative sample of every jurisdiction, negotiation stage or confidential enterprise agreement. Compare its task definition with the buyer’s intended workflow. A benchmark can help establish an annotation vocabulary and a reproducible starting point. It cannot establish rights to unrelated contracts or prove that a private source contains the same labels. Review the original publisher artifacts, instructions and applicable terms. Where a commercial project requires other document types or lawyer-reviewed rationale, those should be named as additional deliverables rather than assumed to come with any corpus bearing the word contracts. Reference What it supplies What it does not establish [Atticus CUAD](https://www.atticusprojectai.org/cuad/) Public clause-review benchmark with labeled agreements Coverage and terms apply to that source, not unrelated private contracts. [LXT document datasets](https://www.lxt.ai/datasets/document/) Advertises custom legal/document annotation work A service capability is not evidence of an existing licensed agreement archive. [Appen enterprise data](https://www.appen.com/enterprise-data) Markets permissioned enterprise documentation workflows No contract-specific price or Fiund supplier relationship is established. Clause labels must describe meaning and boundaries A clause taxonomy needs definitions, inclusion rules and boundary conventions. A provision may span several paragraphs, contain multiple concepts or depend on an exception elsewhere. An annotation can be textually precise while semantically incomplete. State whether labels apply to the entire provision, the operative sentence or a minimal answer span, and explain how repeated or overlapping concepts are represented. Negative examples deserve equal attention. A contract may omit a provision, express it indirectly or place it in a schedule. Absence of a keyword is not proof of absence of the concept. Include uncertainty and disagreement where reviewers cannot resolve the meaning from the supplied material. A useful dataset does not turn every ambiguous case into a false binary answer merely to simplify model training. The annotation record should make the evidence and limits inspectable. Version and amendment lineage protect the task An amendment may replace a provision without restating the complete agreement. A renewal or order form may refer back to terms signed earlier. Record those relationships and any effective dates needed for the task. If a dataset contains several versions of one agreement, keep them linked rather than counting them as unrelated examples of legal diversity. This lineage matters for both model input and evaluation. A question about current obligations may require a different document bundle from a question about the original signed terms. A reviewer needs to know what information was available at the relevant time. Synthetic combinations of clauses should be marked as synthetic, and redacted versions should remain traceable to the approved source record. A neat file hierarchy cannot substitute for an explicit account of which document governs which part of the example. Review rationale is a separate annotation product A reviewer may identify a clause, explain its effect, compare it with a policy or recommend negotiation. Those are different labels. If the buyer needs rationale, define whether it describes textual interpretation, internal business policy or a broader risk judgment. The same wording can produce different recommendations for different organizations or commercial situations. Record the reviewer role, guideline version and adjudication process at the appropriate disclosure level. Multiple reviewers may disagree for principled reasons. Preserve that disagreement where it is useful instead of presenting one judgment as universal legal truth. A source that offers contract text does not automatically offer expert annotations or authority to publish reviewer identities. Keep professional review, commercial use permissions and ordinary text-processing preparation as separate scope questions. Avoid boilerplate and agreement-family leakage Contract corpora often contain repeated templates, standard exhibits and near-identical negotiated versions. A random document split can allow a model to recognize familiar wording rather than generalize to new agreements. Group related families, parties or templates according to the evaluation objective. Where the target task involves amendments, keep the relevant family together unless the protocol deliberately tests temporal updates. Metrics should match the task. Exact span agreement, clause presence, answer correctness and evidence grounding measure different behavior. A model can find the right paragraph while giving an unsupported explanation. Include difficult negatives and cases that require abstention because the relevant schedule or amendment is absent. Report the scope of the test rather than allowing one aggregate score to imply competence across every contract type, jurisdiction or legal review decision. Commercial rights and confidentiality require a bounded source review Publicly filed contracts, licensed archives, internal agreements and synthetic examples have different provenance. Possession of a PDF or access through a public website does not itself settle the proposed model use, redistribution or confidentiality obligations. Ask who can authorize the selected material and which components are included. Attachments, third-party schedules and personal information may need separate treatment. Redaction can affect the task. Removing names may be compatible with clause classification but can break a party-resolution or obligation-linking example. Describe whether identifiers are removed, replaced or consistently pseudonymized, and preserve relationships needed for the model objective. A first sourcing brief should explain requirements without transmitting confidential agreements. Agree appropriate access and handling before a sample exchange, and avoid using private customer names or document extracts as promotional evidence. Pricing and deals: compare the actual deliverable CUAD supplies a public licensed research starting point, not a market average for private contract corpora. The checked sources do not establish a reliable contract-data price per agreement or a disclosed purchase value for a directly comparable private collection. A company acquisition, legal AI investment or software subscription would not be a substitute for such evidence. A useful quote separates source rights, agreement-family reconciliation, text preparation, redaction and expert labels. The number of agreements alone conceals length and complexity: a short standard form and a heavily amended multi-document relationship can require different work. Define the accepted unit, review depth, update scope and allowed model uses. If a provider supplies an estimate, retain its date and assumptions. Fiund’s project discussion should establish these terms for the proposed source rather than applying a speculative rate to every PDF. Example schema: an agreement family with evidence-linked labels A contract delivery can remain compact while preserving the relationships needed for review. The schema below is a design example. It separates source identity, governing context, clause boundaries and interpretation. It does not imply that every field is present in a proposed archive or that a machine-generated label has received professional approval. The most useful identifiers survive redaction and format conversion. A reviewer should be able to find the provision in the approved source version and determine whether an amendment changes the answer. If that connection cannot be made, the collection may still support a limited text task, but should not be presented as complete agreement-family reasoning. State the intended level before paying for unnecessary or unverifiable annotations. Illustrative field Why it matters Acceptance question Agreement family Groups the main agreement, schedules, amendments and related orders. Keep versions together for split design and time-specific questions. Document status Distinguishes draft, executed, superseded and other known source states. Do not infer execution merely from a filename or signature-shaped image. Provision span Connects the label to source text and page coordinates if required. Record overlapping, repeated or distributed provisions explicitly. Label definition Names the taxonomy and guideline version used by reviewers. A familiar clause name may hide materially different inclusion rules. Interpretive context Identifies relevant parties, dates, exceptions and cross-references. Missing referenced material should remain an acknowledged evidence gap. Review record Preserves reviewer role, uncertainty and adjudication method. Professional judgment should not be confused with a keyword-based extraction. Case study design: answer changes after an amendment Prepare an example with an original term and a later amendment affecting the same topic. Ask one question about the original agreement and another about the terms at a later effective date. The expected answers should identify both the applicable source and the reason the other version is not controlling for that question. Include a case where the referenced schedule is absent. The appropriate output may be an evidence limitation rather than a guessed conclusion. This contrast tests whether a dataset supports version-aware reasoning or only recognition of familiar clauses. It also shows why the agreed source bundle matters commercially: several unrelated PDFs are not necessarily equivalent to a coherent, authorized agreement family with reviewable labels. Build a representative contract acceptance sample Choose examples from different agreement families and include amendments, missing references, complex exceptions and negative cases. Review the original text together with labels and evidence spans. A small clean sample selected only for easy clauses cannot demonstrate the quality of a broader negotiation archive. Record both technical preparation defects and unresolved interpretation questions. For a Fiund inquiry, describe the contract types, jurisdictions, dates, intended task, required labels and permitted uses. State whether the project needs text-only material, page images, document relationships or expert rationale. Explain what an acceptable sample would prove and which uncertainty should remain visible. Availability and commercial scope are established for the actual collection. The goal is a defensible data specification that connects source documents to the model’s output, without implying that a category page supplies legal approval or universal contract expertise. Does a contract corpus automatically include expert risk judgments? No. Contract text, clause spans and expert recommendations are separate deliverables. A risk judgment may depend on an organization’s policy and commercial context as well as the wording. Specify the reviewer qualifications, guidelines, evidence and uncertainty needed. Do not treat a labeled clause as a universal recommendation or a substitute for professional review. Can public contracts be mixed with confidential agreements? Only within an appropriately reviewed source and handling scope. Preserve provenance and permission differences, and define how related documents, sensitive fields and redistribution conditions are managed. Public availability of one component does not resolve the status of another. A model-training specification should identify the actual permitted uses rather than assume every PDF in a folder shares one license. How should contract dataset size be reported? Report independent agreement families, document types, versions and useful labeled units where available. Repeated templates and alternate exports can inflate file counts without adding comparable variation. The relevant measure depends on the task: clause examples, complete agreement bundles and time-specific obligation cases are different units. A source manifest should preserve those distinctions so buyers can compare scope. Sources Atticus CUAD: https://www.atticusprojectai.org/cuad/ LXT document datasets: https://www.lxt.ai/datasets/document/ Appen enterprise data: https://www.appen.com/enterprise-data