Contract redline and negotiation datasets for legal AI

A contract redline dataset is the negotiation history of real agreements: each draft exchanged between the parties, the tracked changes and margin comments on every turn, the internal approvals and playbook positions behind them, and the version that was finally signed. SourceX sources these from legal teams and law firms, typically covering several matters or fiscal years, with party names and deal terms removed and privileged material excluded before delivery.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Contract version chains with tracked changes, comments, approvals and executed versions
Typical systems: Document management (iManage, NetDocuments), CLM platforms, shared drives
Typical history: Several matters or fiscal years; varies by partner
Modality: Word documents and PDFs with extracted text and structured change data
Delivery formats: Agreed per order; original DOCX plus JSONL change and clause records
Preparation: Party names and deal terms removed; privileged material excluded; authorship confirmed
Licensing: Rights to client work product and counterparty text confirmed before delivery
Availability: Depends on partners holding complete version chains they can license; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
matter_id	string	Pseudonymous ID for one negotiation, linking every version, comment and approval.
contract_type	enum	Agreement family, such as NDA, master services agreement, SaaS subscription, data processing addendum or lease.
paper	enum	Whose template the negotiation started from, our paper or the counterparty's paper.
party_roles	object	Each side's role, such as customer or vendor, and which side the source partner represents.
versions	array	Every draft in order, with sender side, timestamp, file hash and extracted text.
changes	array	Tracked insertions, deletions and moves per version, mapped to the clause they touch.
comments	array	Margin comments with author side, anchored text and replies, with internal-only notes flagged.
clauses	array	Clause segments with type labels, such as limitation of liability or indemnity, linked to the partner's clause library where one exists.
transmittals	array	Cover emails exchanged with each draft, de-identified; internal emails giving legal advice are excluded.
playbook_position	object	The partner's standard, fallback and walk-away positions for each negotiated clause, where a playbook exists.
approvals	array	Internal escalations and sign-offs, with role, clause, decision and timestamp.
executed_version	object	The signed text and its differences from the last circulated draft.
outcome	object	Signed, abandoned or still open, with elapsed days and number of turns.
redaction	object	Entity types and deal terms masked and the placeholder scheme used.

Example record

{
  "matter_id": "mtr_4b7e21",
  "contract_type": "saas_subscription",
  "paper": "counterparty",
  "party_roles": { "partner_side": "customer", "counterparty_side": "vendor", "counterparty": "[VENDOR_1]" },
  "versions": [
    { "v": 1, "from": "counterparty", "t": "2022-11-03", "sha256": "3f9c…" },
    { "v": 2, "from": "partner", "t": "2022-11-10", "sha256": "a17e…" },
    { "v": 3, "from": "counterparty", "t": "2022-11-21", "sha256": "c402…" }
  ],
  "changes": [
    { "v": 2, "clause": "limitation_of_liability", "op": "replace",
      "from": "fees paid in the three (3) months preceding the claim",
      "to": "fees paid or payable in the twelve (12) months preceding the claim" },
    { "v": 2, "clause": "limitation_of_liability", "op": "insert",
      "text": "except for breach of Section [SECTION_REF] (Data Protection), which is capped at [MULTIPLE] times such fees" },
    { "v": 3, "clause": "limitation_of_liability", "op": "replace",
      "from": "[MULTIPLE] times", "to": "[MULTIPLE_2] times" }
  ],
  "comments": [
    { "v": 2, "side": "partner", "internal": false, "anchor": "limitation_of_liability",
      "text": "Three months does not reflect the data risk. Our standard is 12 months with a data super-cap." },
    { "v": 3, "side": "counterparty", "internal": false,
      "text": "12 months accepted. Super-cap reduced, consistent with our insurance coverage." }
  ],
  "playbook_position": { "clause": "limitation_of_liability", "standard": "12m_fees_plus_data_supercap",
                         "fallback": "12m_fees", "walk_away": "below_6m_fees" },
  "approvals": [
    { "t": "2022-11-22", "role": "deputy_gc", "clause": "limitation_of_liability",
      "decision": "approve_reduced_supercap" }
  ],
  "executed_version": { "v": 4, "t": "2022-12-01", "diff_from_last_draft": "signature_blocks_only" },
  "outcome": { "status": "signed", "turns": 4, "days": 28 },
  "redaction": { "masked": ["PARTY", "PERSON", "PRODUCT", "AMOUNT", "MULTIPLE", "SECTION_REF"] }
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train redlining and drafting assistants

Real turns show what experienced lawyers change in a counterparty's draft, how they phrase a counter-position and which clauses they leave alone.

Model negotiation outcomes

Version chains that end in a signed agreement show which positions held, which were traded away and how many turns each clause took to settle.

Build playbook-compliance review

Clauses paired with a playbook's standard and fallback positions, plus the approvals granted for deviations, label what needed escalation.

Improve clause extraction and classification

Labeled clause segments across many paper sources and drafting styles cover the variation a clause extractor meets in practice.

Create held-out legal evaluation sets

A counterparty draft and the partner's actual response form a graded task: propose a markup and compare it with what a lawyer sent, on documents that were never published.

Use-case guides: Legal AI, Document AI, enterprise search and RAG, Private evaluation sets

What makes this data valuable

Complete version chains

Every turn from first draft to signature, not just the first and last versions.

Both sides' changes

Counterparty markups as well as the partner's own, so a model sees positions argued from each side.

Playbook alignment

Clauses tied to the partner's documented standard, fallback and walk-away positions.

Approval records

Escalations and sign-offs show which deviations required senior authority.

Paper diversity

Negotiations on many counterparties' templates, not only on the partner's own paper.

Clean change data

Tracked changes extracted per version, with handling for moves and for drafts sent without tracking.

Negotiation lives in the drafts

Executed contracts are the most visible legal data: public filings, template libraries and clause banks all hold final versions. A signed agreement shows where a negotiation ended, but it hides the work a legal AI is asked to do: reading the other side's paper, deciding which clauses are acceptable, writing a counter-position that will survive the next turn, and knowing when to escalate instead of conceding. That work exists only in the chain of drafts and in the comments and approvals around them.

The useful unit is the turn: one side's draft, the changes the other side made, the comment that explains them, and what came back. A collection of turns teaches positions and counter-positions in context. The same clause pulls different responses depending on the contract type, whose paper it is, the deal size and the stage of the negotiation, and only histories capture that dependence.

Where version chains break

Missing turns. Drafts often travel by email outside the document management system, and some versions arrive as clean copies without tracked changes. Rebuilding the sequence means comparing documents, and gaps should be marked rather than smoothed over.
Mixed authorship. One file can hold edits from both sides and several internal reviewers, so author metadata is often unreliable. Side attribution by turn is more trustworthy than by edit.
Template echoes. The partner's own paper appears in every negotiation it started. Without deduplication, a model learns one company's template wording as if it were market standard.
Jurisdiction skew. Governing law, contract types and industries tend to concentrate around a partner's practice, so check coverage before treating the data as representative.

Rights questions come first. Before anything is licensed, SourceX confirms who owns each document set and agrees how counterparty text will be treated. The rest of a request is usually framed by contract type, governing law, the side the partner represents, how complete the version chains are and whether a playbook exists to align clauses against.

What to check before licensing

Confirm who owns the documents. A law firm's drafts for clients are client matters, and the firm may need each client's informed consent before reuse, under professional confidentiality rules for lawyers.
Ask how privileged material was identified and excluded, including internal comments to counsel, legal memos stored with the contract and email cover notes giving legal advice.
Check confidentiality clauses in the agreements themselves. Many contracts make their own terms confidential, and that obligation can extend to the drafts exchanged during negotiation.
Agree how counterparty-authored text is treated. Their template and markups were written by another organization, which never agreed to the partner licensing them.
Search the sample for identifiers that de-identification often misses: defined terms built from party names, product names, notice addresses, signature blocks and deal-specific numbers.
Check version completeness: whether gaps exist where drafts were exchanged by email only, and how versions sent without tracked changes were compared.
Check coverage by contract type, governing law and whose paper each negotiation started on. A set dominated by one template or one jurisdiction will not generalize.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can law firms license client contracts and redlines for AI training?

Only with authority from the clients concerned. Contracts, drafts and redlines a firm prepares are client matters, and lawyers' professional rules generally prohibit revealing information relating to a client's representation without the client's informed consent. A firm that participates needs that consent for the matters in scope. In-house legal teams licensing their own company's agreements face fewer layers of approval.

How is attorney-client privilege protected?

Privileged material is excluded rather than redacted in place. That covers internal comments seeking or giving legal advice, legal memos filed with the contract and cover emails with advice. Text exchanged with the counterparty is usually not privileged because it was shared with the other side, but it can still be confidential under the agreement or a non-disclosure agreement, which is a separate question.

What happens to the counterparty's text in a redline?

It is de-identified like the partner's text, and its use is part of the rights review. The counterparty's template and markups appear in every version, so excluding them would leave only one side of the negotiation. Whether they can be included, and in what form, is agreed per order based on what the partner's agreements and confidentiality obligations allow.

How are deal terms like price and liability caps de-identified?

Commercially sensitive values are replaced with typed placeholders or ranges as agreed, so a model still sees that a cap changed from a three-month to a twelve-month fee basis even when the multiple itself is masked. Party, product and person names become placeholders, including inside defined terms. Which values stay visible is set during scoping.

How is this different from public contracts such as SEC EDGAR exhibits?

EDGAR exhibits show where a negotiation ended; redline histories show how it got there. Exhibits are final agreements, or forms of them, that public companies filed, mostly because they were material, and filers may redact terms that are not material and that they treat as private or confidential. Public research sets such as CUAD are drawn from them. Redline histories add the intermediate drafts, both sides' changes, comments and approvals, from negotiations that were never filed.

Do redline histories include the reasoning behind each change?

Partly. Margin comments exchanged with the other side explain many changes, and approval records explain internal escalations. Much of the reasoning sat in calls or privileged internal discussion that is excluded, so the dataset shows what changed and when more reliably than why. Playbook positions help fill that gap where the partner has a playbook.

Enterprise document archives
A company's working files with folders, versions and sharing metadata
SOPs, playbooks and internal knowledge bases
Written procedures with page history, ownership and links to execution records
Human feedback and QA-scored work
Work items with scores, verdicts and corrections from the people who reviewed them
Approved workplace email and chat exports
Approved, de-identified exports of team email threads and chat channels
Enterprise workflow and task execution histories
Linked task trajectories from request to outcome, across every tool the work touched

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.