TenantForge sample tenant: evaluation licence. By requesting access you agree to use this synthetic dataset only for internal evaluation, testing and benchmarking; not to redistribute it; not to represent any of its content as real or use its synthetic identities to create or seed real accounts; and not to use it to train a product or service that is sold without a commercial licence from Tidewater Labs (legal@tidewaterlabs.tech). Full text in LICENSE.

TenantForge sample tenant: Tidal Bend Savings Bank

A synthetic banking enterprise tenant (GLBA, PCI DSS, SOX, FCRA) for testing DLP and DSPM tools and evaluating AI agents on sensitive data: 186 files (Word, Excel, CSV, PDF, scanned PDF, PNG, text, Markdown, .eml, mbox, chat exports, code, bucket listings, zips), identities, permissions, a year of activity logs, and an exact label for every one of 764 sensitive values plus 268 enumerated decoys. No real personal data by construction. Same inputs rebuild byte for byte on any operating system.

tenantforge-sample-banking.zip: the full tenant. Unzip and run python validate.py . to re-check every label against the content.
rows.jsonl: 167 text-bearing artifacts flattened for the dataset viewer (text plus findings with locators). Image-only scans and zip members are only in the archive.
labels.jsonl, artifacts.jsonl, manifest.json: the label set, artifact inventory and provenance, as shipped.
detector.json: the baseline recall numbers behind the leaderboard below.

Tenant id tf_d6c9735a203a, content hash 260fca9873c3cedb1b5fb627a8876ce9b9f2974fa9511e1f18245a6a81138e8c, generator tenantforge 0.1.2.

Other industries in the free sample collection: https://huggingface.co/collections/jeffreyphall/tenantforge-sample-tenants-6ae8f9824aead1f72f5a181f

Licence: evaluation only (see LICENSE). Standard (5,000) and large (50,000) tenants in sixteen industries and regimes are available under a commercial licence: https://tidewaterlabs.tech/tenantforge/

Detector leaderboard

Recall of two free baselines over every labelled true value in the text-bearing artifacts (764 values; 40 image-only and 124 zip-member values excluded because these baselines do not open them). Format regexes are the plain patterns in tenantforge/detector.py; Presidio is Microsoft Presidio 2.2.366 with spaCy en_core_web_sm, score threshold 0.4. Raw numbers in detector.json.

data type	labelled	format regex	Presidio
ACCESS_TOKEN	4	1.00	0.00
API_KEY	6	0.67	0.00
BANK_ACCOUNT	25	0.00	0.40
BOARD_CONFIDENTIAL	4	0.00	0.00
CLOUD_CREDENTIAL	4	0.50	0.00
CONTRACT_TERM	49	0.00	0.00
CREDIT_SCORE	16	0.00	0.00
DB_CONNECTION_STRING	10	1.00	0.00
DOB	30	0.00	0.77
FINANCIAL_PRE_RELEASE	28	0.00	0.00
HOME_ADDRESS	38	0.00	0.71
IBAN	3	0.00	1.00
LEGAL_PRIVILEGED	38	0.00	0.00
MNA_MATERIAL	3	0.00	0.00
PAN	22	0.50	0.41
PAN_LAST4	11	0.00	0.00
PASSWORD	5	0.00	0.00
PERFORMANCE_REVIEW	6	0.00	0.00
PERSONAL_EMAIL	60	0.37	0.37
PERSONAL_PHONE	22	0.46	0.46
PERSON_NAME	212	0.00	0.72
PHI_INSURANCE_ID	7	0.00	0.00
PHI_MRN	8	0.50	0.00
PRIVATE_KEY	3	1.00	0.00
ROUTING_NUMBER	6	0.67	0.00
SALARY	42	0.00	0.00
SECURITY_VULN_REPORT	3	0.00	0.00
SOURCE_CODE	5	0.00	0.00
SSN	82	0.65	0.55
SSN_LAST4	4	0.00	0.00
TAX_ID	8	0.00	0.00
all (weighted)	764	0.17	0.39

Decoys fired on: format regex 118, Presidio 94. A decoy is an enumerated value that looks sensitive and is not (test card numbers, order numbers shaped like SSNs, lorem keys); every fire is a false positive the label set lets you count exactly.

Add your tool. Run it over the unzipped tenant, score it with python validate.py . --compare your-hits.jsonl (one {path, value, data_type} per line) or your own matcher against labels.jsonl, and open a discussion on this repo with the per-type recall and decoy-fire counts. Rows from DLP and DSPM vendors are listed with the product and version named.

Purpose

A synthetic enterprise tenant (files, mail, chat exports, code snippets, cloud bucket listings, identities, permissions and activity logs) with exact ground-truth labels, for testing and benchmarking data-loss-prevention (DLP) and data-security-posture (DSPM) tools and for evaluating AI agents that must handle sensitive data correctly.

Company: Tidal Bend Savings Bank (banking), 1200 staff, regimes: GLBA, PCI_DSS, SOX, FCRA. Timeline 2025-01-06 to 2025-12-19, evaluated as of 2026-01-15.

item	count
artifacts (top level)	186
artifacts incl. attachments and archive members	218
true findings (labels.jsonl, is_decoy=false)	764
decoys (labels.jsonl, is_decoy=true)	268
identities / groups	1224 / 28
folders / buckets	84 / 4
log events / planted scenarios	8459 / 6

format	artifacts
bucket_listing	4
chat_json	4
code	14
csv	13
docx	62
eml	9
mbox	1
md	16
pdf	28
pdf_scanned	13
png	2
txt	20
xlsx	25
zip	7

Files

files/, buckets/, mail/, chat/, repos/   the artifacts
logs/file_access/*.jsonl                 access events by month
logs/sharing_changes.jsonl, downloads.jsonl
logs/scenarios.jsonl                     ground truth for planted exfiltration-like sequences
identities.jsonl groups.jsonl acl.jsonl buckets.jsonl subjects.jsonl
artifacts.jsonl                          one row per artifact (exposure is authoritative here)
labels.jsonl                             one row per labeled occurrence (schema: finding.schema.json)
text_parts.jsonl                         extracted PDF text the offsets refer to
manifest.json                            provenance and file hashes
validate.py                              re-checks every label against the content

Label schema

See schemas/finding.schema.json. Each finding has one or more locators (char_span, cell_ref, json_pointer, image_region, archive_member, email_attachment), a data type, sensitivity tier, applicable regulations for this tenant, the exposure context, decoy flag and hard-case tags. Decoys are values that look sensitive and are not; they are enumerated so false positives can be scored.

Generation method

Everything is a pure function of (generator version 0.1.2, config hash 63e5ce70f5af7637..., seed 20260207). The structure layer (identities, share tree, permissions, which values go where) is deterministic Python. Values are injected by that layer, so each label records the exact location rather than being inferred from text afterwards. Prose comes from rule-based templates. Local model used for text: qwen3:14b revision ollama:bdbd181c33f2ed1b31c972991882db3cf4d192569092138a7d29e973cd9debe8, licence Apache-2.0, source hf:Qwen/Qwen3-14B. No text from any commercial LLM API is included.

Content hash (all files except manifest.json): 996688bd4923e4011c78937b94eee7aef6f373e5f851e520b5d7753de9c646b5. Dependency versions are recorded in manifest.json; rebuilding with the same versions reproduces every byte.

Safety construction (no real personal data)

SSNs use area 900-999 or 000 with group digits outside the ITIN ranges, so no value can be an issued SSN or ITIN. Detectors that validate SSN area numbers will therefore reject every true SSN here; score them in format-only mode.
Payment card numbers are the card networks' and processors' published test numbers only (37 distinct values; expect repetition). Decoys flip one digit so Luhn fails.
Phone numbers are 555-0100 to 555-0199. Email addresses and domains use example.com/org/net or reserved .example/.test names. Routing numbers use unassigned ABA prefixes 13-20 with a valid check digit. EINs use the unassigned 00- prefix. IBANs are published examples.
Names are seeded Faker output screened against a public-figure list. They are common names and will coincide with private individuals who share them; that is unavoidable and is not personal data about those individuals.
Addresses are synthetic street names with 5-digit house numbers and random ZIP codes that do not correspond to the city.
Driver's licence and passport numbers are format-plausible only and are disabled by default; do not rely on them being unissued.
Private keys are PEM-shaped random data, not valid keys.

Known limitations

Prose is templated and repetitive; it is not a realism benchmark for text.
The scanned-document and screenshot artifacts are rendered, not photographed.
Activity logs are statistically shaped, not replayed from a real system.
Only en_US names and US geography are generated in this version.
Pre-release financial figures, contract terms and similar business-confidential content are invented; nothing refers to a real company.

Intended uses

Detector precision/recall measurement, DSPM exposure-finding evaluation, AI-agent red-teaming on sensitive-data handling, training data for classification models where synthetic provenance is acceptable, demos.

Prohibited uses

Representing any content here as real; using the synthetic identities to impersonate or to seed real-world accounts; redistribution outside the terms in LICENSE.

Licence

This tenant ships under the evaluation licence text in LICENSE (evaluation: free sample, internal evaluation and benchmarking only; commercial: per-organisation internal use including training; placeholder: no rights granted beyond evaluation on the recipient's own systems).