Data Corpus Schema Documentation
Version: 1.0
Overview
Each record in this corpus represents a single annotated case file — a real-world professional communication thread, enriched with structured metadata and associated document artifacts. Records are JSON objects stored as elements within a top-level array.
The corpus is intended for post-training of large language models on professional, domain-specific reasoning tasks — particularly legal drafting, document review, multi-turn client-lawyer dialogue, and expert question-answering.
Top-Level Record Structure
Record (Object)
├── trace_id string Unique identifier for the record
├── metadata object Classification and provenance signals
└── data object The core content payload
Field: trace_id
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Format |
Alphanumeric string (e.g., "fd1748286", "era1711548", "cqa153562872") |
| Description |
Globally unique identifier for this record. Used for deduplication, lineage tracking, and cross-referencing attachments. Attachment file_path values typically embed this ID (e.g., /document/SAFEv1-trace_id-fd1748287). |
metadata Object
Carries classification and provenance signals consumed by data pipelines and model trainers for filtering, weighting, and curriculum design.
metadata (Object)
├── content_type string Content format identifier
├── area_of_law string Legal domain
├── document_type string[] Legal instrument type(s) [OPTIONAL — see note]
├── industry string[] Industry vertical(s)
├── jurisdiction string[] ISO 3166-1 alpha-2 jurisdiction code(s)
├── pii_redacted boolean PII redaction status
├── proprietary_status string[] IP/ownership classification(s)
├── provenance_tier string Source quality tier
├── tokens_estimate integer Approximate token count
└── language string BCP 47 language tag
Field: metadata.content_type
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Example |
"communication-thread-with-documents" |
| Description |
Identifies the structural format of the record. Governs which fields are expected under data, and in particular which context sub-schema applies. |
Known values:
| Value |
Description |
data.context required? |
communication-thread-with-documents |
Multi-turn message thread with one or more document attachments and rich structured context |
Yes |
final-document |
A finished legal instrument produced through an internal drafting workflow; thread captures internal handoff, not client dialogue |
Yes |
expert-reasoning |
A professional legal analysis or opinion delivered in response to a client question; thread is typically short (1–2 turns) |
Yes |
client-question-expert-answer-pair |
A single client question paired with a single expert answer; may lack a context block |
No |
Field: metadata.area_of_law
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Example |
"corporate" |
| Description |
High-level legal domain of the content. Used for domain-stratified sampling. |
Known values (non-exhaustive): corporate, tax, compliance, ip, employment, litigation, regulatory
Field: metadata.document_type
| Property |
Detail |
| Type |
string[] |
| Required |
Conditional — required for all content_type values except client-question-expert-answer-pair |
| Example |
["SAFE"] |
| Description |
One or more legal instrument types referenced or produced in the thread. Must be an array of strings even when only one type is present. Omit the field entirely (do not use null or []) for records where no document type applies. |
Known values (non-exhaustive):
| Canonical value |
Observed variants |
SAFE |
— |
term-sheet |
term_sheet |
token-warrant |
— |
legal-opinion |
— |
tax-memo |
tax_memo |
SHA |
— |
NDA |
— |
employment-agreement |
— |
board-resolution |
— |
Field: metadata.industry
| Property |
Detail |
| Type |
string[] |
| Required |
Yes |
| Example |
["finance", "web3"] |
| Description |
Industry vertical(s) of the parties or subject matter. A record may span multiple industries. |
Known values (non-exhaustive): finance, web3, technology, real-estate, healthcare
Field: metadata.jurisdiction
| Property |
Detail |
| Type |
string[] |
| Required |
Yes |
| Format |
ISO 3166-1 alpha-2 country codes |
| Example |
["SG"], ["UK", "PT", "PL"] |
| Description |
Governing legal jurisdiction(s). A record may span multiple jurisdictions (e.g., cross-border tax advice or multi-party deals). |
Field: metadata.pii_redacted
| Property |
Detail |
| Type |
boolean |
| Required |
Yes |
| Example |
true |
| Description |
Indicates whether personally identifiable information has been removed. Records with false must not be used in training without further review. |
Redaction markers in content: PII is replaced inline using typed angle-bracket placeholders. All placeholders are uppercase and may be indexed with a numeric suffix.
| Placeholder |
Represents |
Notes |
<NAME> |
Person's name |
May appear as <NAME_1>, <NAME_2> when multiple individuals appear in one record |
<NUMBER_N> |
Numeric value |
Indexed per record (e.g., <NUMBER_1>, <NUMBER_5>); indices are local to each record |
<EMAIL> |
Email address |
— |
<COMPANY> |
Company or entity name |
— |
<ENTITY> |
Named legal entity (e.g., DAO, foundation, protocol) |
Observed in web3/compliance records; distinct from <COMPANY> |
<ENTITY_N> |
Named legal entity (indexed) |
e.g., <ENTITY_2> for a second entity within the same record |
<DATE> |
Specific date |
— |
Field: metadata.proprietary_status
| Property |
Detail |
| Type |
string[] |
| Required |
Yes |
| Example |
["IP_acquisition"] |
| Description |
One or more ownership and licensing classifications for the record. Governs permissible training use. Multiple values may apply simultaneously. |
Known values:
| Value |
Description |
IP_acquisition |
IP rights formally acquired; cleared for training use |
terms_consent |
Source explicitly consented to use under platform terms of service |
licensed |
Used under a separate data license agreement |
public_domain |
No restrictions |
restricted |
Do not use without explicit written clearance |
Field: metadata.provenance_tier
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Example |
"t1" |
| Description |
Source quality tier indicating data origin and reliability. Used for training curriculum weighting. |
Known values:
| Tier |
Description |
t1 |
Verified primary source: practitioner-generated, matter-linked content (e.g., actual client files, legal opinions) |
t2 |
Professional practitioner-generated content not directly tied to a verified client matter |
t3 |
Lightly curated secondary or synthetic material |
Field: metadata.tokens_estimate
| Property |
Detail |
| Type |
integer |
| Required |
Yes |
| Example |
11232 |
| Description |
Estimated token count for the full record, including all nested text fields. Used for batch sizing and context window planning. Counts are approximate and may not include attachment content. |
Field: metadata.language
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Format |
BCP 47 language tag |
| Example |
"en-US" |
| Description |
Language and locale of the record content. |
data Object
The primary content payload. All records contain communication_thread. The context block is present in most records but is optional for client-question-expert-answer-pair content types.
data (Object)
├── context object Structured summary of matter parameters [OPTIONAL]
└── communication_thread object[] Ordered sequence of messages
data.context Object
Provides machine-readable parameters extracted or inferred from the thread. The internal structure of context varies significantly by content_type and document_type. All fields are descriptive summaries, not validated legal data. Numeric fields may contain redaction placeholders and should be treated as strings.
data.communication_thread Array
An ordered array of message objects representing the full dialogue. Thread length varies by content type: client-question-expert-answer-pair records are exactly 2 turns; communication-thread-with-documents records may have 5 or more.
communication_thread (Object[])
└── message_object
├── sequence integer 1-based ordinal position in the thread
├── sender string Role of the message author
├── message string Full message body (may contain redaction placeholders)
└── attachments object[] Documents attached to this message (may be empty array)
Field: sequence
| Property |
Detail |
| Type |
integer |
| Required |
Yes |
| Description |
1-based index. Defines chronological order. Gaps in sequence values are not expected but should be handled gracefully by consumers. |
Field: sender
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Description |
Role identifier of the message author. |
Known values:
| Value |
Description |
Applicable content types |
client |
The instructing party (company, founder, or business client) |
all |
lawyer |
The legal professional advising the client |
all |
user_1 |
Internal team member initiating a drafting task |
final-document |
user_2 |
Internal team member executing a drafting task |
final-document |
Field: message
| Property |
Detail |
| Type |
string |
| Required |
Yes |
| Description |
Full text of the message. May span multiple paragraphs. Redaction placeholders appear inline where PII was removed. Newlines are represented as \n. Informal register and typos may be present in client messages, particularly in client-question-expert-answer-pair records. |
Field: attachments
| Property |
Detail |
| Type |
object[] |
| Required |
Yes |
| Description |
List of documents attached to this message. Must be an empty array [] when no attachments are present; the field must not be omitted. |
Attachment object structure:
attachment (Object)
├── file_name string Human-readable filename including extension
└── file_path string Corpus-internal path to the document record
Observed file_path formats:
Format — document shares parent trace_id, with explicit file extension:
/document/{document_name}-trace_id-{parent_trace_id}.docx
Example: /document/tax_memo-trace_id-fd1748284.docx
/document/term_sheet-trace_id-fd1748285.docx