Skip to content

Document AI data

Contract Families: Master Agreements Linked to Amendments, SOWs and Exhibits

Quick answer

A contract amendment linking dataset is a set of real agreements grouped into families, where each document carries a pointer to its parent agreement, its position in the amendment sequence, its effective date and the clauses it supersedes. These labels let you train and evaluate models that reconstruct current effective terms, a task that clause-level datasets such as CUAD do not cover [2]. Buyers should specify family completeness, link provenance, scan-versus-native mix and consistent party masking before anything else.

By SourceX Editorial · Updated

  1. 01Your data spec
  2. 02Matched to businesses
  3. 03Rights and samples reviewed
  4. 04Licensed delivery
How it works with SourceX: Your data spec → Matched to businesses → Rights and samples reviewed → Licensed delivery.Illustrative

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why clause labels inside one contract cannot teach effective-terms reasoning

Clause labels describe what a single document says, while family labels describe which document controls today. CUAD annotates 510 commercial contracts with more than 13,000 labels across 41 clause categories, but each contract is labeled on its own [2]. ACORD targets clause retrieval for drafting [3], and LegalBench spans 162 reasoning tasks [4]; none of them asks a model to resolve a liability cap that was raised in Amendment 2 and narrowed again in an SOW-specific carve-out.

That resolution is the actual job in a CLM migration or a legal AI assistant. A buyer reconstructing a supplier relationship must answer questions like "what is the current termination-for-convenience notice period?" when the MSA says 90 days, Amendment 1 says 60, and a later restated agreement silently reverts to 90. If you only need per-document clause extraction, start from contract clause annotation data instead; this page covers the relationships between documents.

What a contract family contains in practice

A typical commercial family is a master services agreement (or master purchase, license or subscription agreement) with a long tail of dependent documents. Common members include:

  • Amendments and restatements: numbered or dated changes; an amended and restated agreement resets the base text.
  • Statements of work and order forms: project- or purchase-specific terms that incorporate the master by reference and often override it for one engagement.
  • Exhibits, schedules and addenda: pricing schedules, SLAs, data processing addenda, security exhibits and business associate agreements.
  • Change orders, side letters and waivers: short documents that alter a single term, frequently missed by CLM migrations.
  • Assignment, novation and termination notices: events that change parties or end parts of the family.

The same structure appears in recurring data purchases, where a master license governs individual order forms; see master data license agreements and order forms for that pattern from the buyer's side.

Why public filings give you incomplete families

SEC filings are the obvious public source of linked contracts, but they rarely contain whole families. Material agreements surface on EDGAR as individually numbered exhibits, such as a master data license agreement attached to a Form 10-K as Exhibit 10.36 [1]. Each exhibit is a standalone file, so the master, its amendments and its order forms arrive as separate documents filed at different times, often with schedules omitted or confidential terms redacted.

The practical result: an EDGAR family often has the master and a few amendments, with pricing schedules omitted, figures bracketed out, and SOWs that never crossed the materiality threshold. Exhibit index titles such as "Amendment No. 3 to Supply Agreement" give a weak link signal, but filers do not use a shared identifier scheme. Models trained only on these families learn a cleaner, more complete hierarchy than they will meet in an enterprise repository.

Enterprise repositories have the opposite problem. Older amendments exist only as executed scans or faxes, newer documents are native PDF or DOCX exports from DocuSign or a CLM, and the same agreement appears under three filenames. For the capture side, see degraded scans, faxes and phone photos; for packets where one PDF bundles several family members, see page-stream segmentation data.

The label set that makes a family usable

A usable family dataset labels the graph, the timeline and the clause-level supersession, not just the documents. The minimum label layers are:

LayerLabelTypical failure if missing
Graphfamily_id, doc_id, parent_doc_id, relation_type (amends, restates, incorporates, attaches_to, terminates)SOWs linked to the wrong master when a counterparty has two MSAs
Sequencesequence_no, execution_date, effective_date, expiration_dateOrdering by execution date when an amendment is retroactive
Supersessionaffected_section, operation (replace, insert, delete, conditional), superseded_byModel treats deleted language as still in force
PrecedenceOrder-of-precedence clause reference and resolved rankingSOW terms override the MSA even where the MSA says it controls
Effective termsResolved value per term key as of a stated dateNo ground truth for evaluation
ProvenanceLink method (explicit reference, metadata, human judgment), annotator, confidenceInferred links cannot be audited

Ask for the precedence ranking explicitly. Many MSAs state that the master controls over any SOW "unless the SOW expressly references the section it modifies," and a model that ignores that condition will produce plausible but wrong effective terms.

Illustrative family record schema

A JSON Lines layout keeps one document per line and makes families easy to stream and split [7]. Describe the files and fields in Croissant so loaders and reviewers read the same structure [6].

Illustrative example: invented to show structure; it does not describe an available dataset.

{"family_id":"F-0412","doc_id":"F-0412-D00","doc_type":"msa","parent_doc_id":null,"relation_type":null,"sequence_no":0,"execution_date":"2019-03-14","effective_date":"2019-04-01","source_format":"scan_tiff","ocr_engine":"recorded","party_tokens":["[PARTY_A]","[PARTY_B]"],"precedence_rank":1}
{"family_id":"F-0412","doc_id":"F-0412-D01","doc_type":"amendment","parent_doc_id":"F-0412-D00","relation_type":"amends","sequence_no":1,"execution_date":"2020-11-02","effective_date":"2020-10-01","retroactive":true,"supersessions":[{"affected_section":"9.2","operation":"replace","term_key":"liability_cap","new_value":"2x annual fees"}]}
{"family_id":"F-0412","doc_id":"F-0412-D02","doc_type":"sow","parent_doc_id":"F-0412-D00","relation_type":"incorporates","sequence_no":2,"effective_date":"2021-02-01","overrides_master":["12.1"],"override_permitted_by":"MSA 1.3"}
{"family_id":"F-0412","eval_item":"Q-0412-07","as_of":"2022-01-01","question":"What liability cap applies to SOW 2 work?","answer":"2x annual fees","evidence_docs":["F-0412-D01","F-0412-D02"],"evidence_pages":[[2],[1,4]]}

The last line shows the most valuable output of a labeled family: an effective-terms question with an as_of date, a resolved answer and page-level evidence across more than one document.

Effective-terms questions as an evaluation set

Effective-terms questions make strong evaluation items because the answer depends on reading several documents in the right order. Public long-document benchmarks show the format: MMLongBench-Doc builds 1,062 expert-annotated questions on 130 lengthy PDFs and annotates the evidence pages and source types behind each answer [5]. A contract-family eval adds two things those benchmarks lack: a temporal as_of parameter and a precedence rule.

Good item types to request:

  • Point-in-time lookup: the value of one term on a given date, crossing at least one amendment.
  • Supersession trap: a question whose naive answer comes from deleted or replaced text.
  • Retroactivity: an amendment executed after the date asked about but effective before it.
  • SOW override: a term changed for one engagement only, with the master unchanged elsewhere.
  • Unanswerable: a family where a referenced exhibit is missing, so the correct output is "not determinable."

Hold whole families out of training, not individual documents, or near-duplicate boilerplate across siblings will leak answers. For broader formats, see document question answering data and multi-page and long document data.

Party masking and confidentiality across a whole family

De-identification has to be consistent across every document in a family, or the links break. If "Acme Logistics" becomes [PARTY_B] in the MSA but [ORG_17] in Amendment 3, a linking model loses its strongest signal and an evaluation answer may become unverifiable. Ask suppliers to apply one pseudonym map per family, including signature blocks, notice addresses, letterhead and scanned stamps that OCR may miss.

Names of signatories, emails, phone numbers and bank details in payment schedules need the same treatment. Commercial terms are a separate issue from personal data: the counterparty's pricing and confidentiality clauses may restrict reuse, and the FTC has warned that companies can face liability when they break promises not to use customer data for undisclosed purposes such as model training [9]. Buyers should see who approved the release and on what basis.

Buyer checklist for a contract family request

Write the request around families and relationships, not document counts. The document dataset requirements spec covers general fields; add these family-specific items:

  1. Family definition: what counts as one family (master plus all dependents, or one SOW chain), and how restated agreements are handled.
  2. Completeness flag: per family, whether all referenced exhibits and amendments are present, with a list of missing references.
  3. Link provenance: whether links come from CLM metadata, explicit text references or annotator judgment, with confidence values.
  4. Format mix: share of scanned versus native documents, OCR engine and version, and whether redlines or only executed copies are included.
  5. Date semantics: separate execution, effective and expiration dates, with retroactive amendments flagged.
  6. Masking map: one consistent pseudonym scheme per family, the method recorded, and a sampled check.
  7. Eval split: family-level held-out set with as_of questions and evidence pages.
  8. Documentation: a Data Card describing sources, annotation method and known gaps [8].

Negotiated drafts and markup history are a different asset; SourceX covers that on its contract redline datasets page.

How SourceX approaches contract family requests

SourceX sources operational datasets, including documents and legal workflows, from US companies, and manages the licensing and ongoing purchases for AI teams wherever they are based. Data is sourced on request rather than held in stock, so a described family structure is a search brief, not inventory, and a request does not guarantee a match. Every release is approved by the supplying company, each dataset is rights-reviewed for ownership and consents, and personal details such as names, emails, phone numbers and account numbers are removed or replaced before delivery, with the method recorded and a sample checked; no method is perfect. You can describe the contract families you need without naming specific businesses.

For the wider map of document tasks, return to the Document AI data hub.

SourceX runs a Find, Assess, Agree, Transact and Manage process, and nothing is contracted until a supplier agrees. Each dataset is delivered under a license that defines the records, allowed uses, term and delivery. Describe your contract family requirements to SourceX.

Sources

  1. U.S. Securities and Exchange Commission, EDGAR, "EX-10.36 Document (Form 10-K exhibit, fiscal 2023)" (2024). https://www.sec.gov/Archives/edgar/data/0001758730/000175873024000025/twm-2023x12x31xexx1036.htm
  2. Hendrycks et al. (The Atticus Project), arXiv, "CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review" (2021). https://arxiv.org/pdf/2103.06268
  3. arXiv, "ACORD: An Expert-Annotated Retrieval Dataset for Legal Contract Drafting" (2025). https://arxiv.org/pdf/2501.06582
  4. Hazy Research, Stanford University, "LegalBench: A collaboratively built large language model benchmark for legal reasoning". https://hazyresearch.stanford.edu/legalbench
  5. Ma et al., arXiv (NeurIPS 2024), "MMLongBench-Doc: Benchmarking Long-context Document Understanding with Visualizations" (2024). https://arxiv.org/abs/2407.01523
  6. Akhtar et al. (MLCommons), arXiv, "Croissant: A Metadata Format for ML-Ready Datasets" (2024). https://arxiv.org/pdf/2403.19546
  7. jsonlines.org, "JSON Lines". https://jsonlines.org/
  8. Pushkarna, Zaldivar, Kjartansson (Google Research), arXiv, "Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI" (2022). https://arxiv.org/pdf/2204.01075
  9. Federal Trade Commission, Office of Technology, "AI Companies: Uphold Your Privacy and Confidentiality Commitments" (2024). https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/01/ai-companies-uphold-your-privacy-confidentiality-commitments

Tell us what your models need

Share scope, volume, language, format, timing and licensing requirements.

Request data