pbhappliedsystems/veritruct-cloud-regulated-deid-1k

Veritruct Cloud — Construction-True Regulated-Domain De-Identification Dataset

Run run_cloud_20260706_000447 · Flagship release demo dataset · PBH Applied Systems, LLC

Generated, gated, masked, labeled, and evaluated by PBH Applied Systems, LLC — Applied AI/ML Consulting · Quality-Gated Synthetic Data · LLM Optimization & Deployment

📄 Read the whitepaper: Veritruct: Quality-Gated Synthetic Data Generation for Regulated Industries, or for a quick overview, read the blog Quality-Gated Synthetic Data for Regulated Industries.

📋 This is a deliberately over-documented dataset card, and that is the point. The synthetic-data ecosystem on the Hugging Face Hub ships datasets with, in the words of the accompanying whitepaper, "row counts, a license, and little else" — the generation pipeline opaque, the rejection criteria unstated, the compliance coverage unaudited. De-identification corpora carry a further, subtler problem: the span labels used to train and benchmark the models that redact clinical and legal records are frequently unverifiable, because a corpus labeled by the same class of tool it is meant to evaluate cannot establish ground truth. This dataset is the deliberate inverse on both counts. Its labels are correct by construction — the personal-data values are injected into the text at offsets the pipeline itself records — and every number below is read directly from the machine-readable artifacts shipped alongside the data (run_manifest.json, evaluation_report.json, resolved_config.yaml). The methodology is documented in full in the whitepaper, Veritruct: Quality-Gated Synthetic Data Generation for Regulated Industries (Hill, 2026).

---

TL;DR

Records (accepted, delivered)	1,053
Records rejected (logged, not delivered)	197
Total attempted	1,250
Overall yield	84.2%
Record type	Construction-true de-identification — {text, spans, masked_text}
Domains	7 — privacy, GDPR, medical, education, transport, ecommerce, legal
Ground-truth spans	2,452 across 17 identifier types (labels correct by construction)
Detector micro-averaged F1 (hybrid regex + GLiNER2)	0.905
Injected-value validity	2,246 well-formed / 206 not (91.6%)
Substrate	Veritruct Cloud — Modal A10G + vLLM
Generation model	Mistral-Nemo-Instruct-2407 AWQ (in-house)
Reconciliation status	ok — all counts and the seven-domain distribution agree

This is the all-domain Cloud de-identification run documented in the whitepaper (§8.5), the counterpart to the single-domain medical run — it exercises the full identifier vocabulary rather than the code-heavy medical subset.

---

What This Dataset Is

This is a construction-true synthetic de-identification dataset spanning seven regulated domains — the training and benchmarking substrate a PII/PHI-redaction model requires: text with known, labeled personal-data locations. It is precisely what cannot be safely drawn from real records, because real records carry real identifiers.

Each record is a completed professional document (a DSAR log, breach notification, clinical note, billing audit, litigation document, ALPR oversight finding, learning-analytics record, and so on) paired with span-level labels marking the exact character offset, length, and type of every personal-data element the text contains, plus a fully anonymized view of the same document.

The defensible core is how the labels are produced. A conventional synthetic de-id corpus is generated first and then labeled by a tagging pass — a regex sweep or a model — so the labels are only as good as the tagger and inherit exactly the errors a de-id benchmark exists to measure. Veritruct inverts this: the labels are not applied after the fact. The personal-data values are injected into the text at positions the pipeline itself chooses and records, so the ground-truth span set is a byproduct of construction rather than the output of a fallible post-hoc tagging step. This is what "construction-true" means, and it is what lets this dataset serve as ground truth rather than as another fallibly-tagged corpus.

Every delivered record also cleared the shared Veritruct quality cascade — a dual-signal hallucination gate and a template-leak gate (which rejects fill-in-the-blank skeleton output rather than completed records). Records that did not clear the gates are written to rejected.jsonl and summarized in rejection_audit.json.

This release was generated on Veritruct Cloud, the cloud-GPU substrate — generation offloaded to a Modal A10G served by vLLM. Its sibling, Veritruct Studio, runs the identical shared core locally on a consumer RTX 4090; only the generation backend differs. A single whitepaper documents the whole system.

---

Dataset Structure

Files

File	Contents
README.md	This dataset card.
config_hash.txt	The SHA-256 checksum of resolved_config.yaml (23f23eb0…4bf74) — identical to the value stamped in run_manifest.json, so the shipped configuration can be verified byte-for-byte against the one the run recorded.
evaluation_report.json	The machine-readable detector evaluation — per-type and averaged precision/recall/F1 of a hybrid regex + GLiNER2 detector scored against the construction-true ground-truth spans, plus the injected-value validity tally.
output.jsonl	The 1,053 accepted, delivered records. One JSON object per line.
rejected.jsonl	The 197 rejected records, each carrying the gate flags that caused rejection.
rejection_audit.json	A roll-up rejection audit: a rejection_summary (total rejected, per-reason counts, per-domain counts) plus all 197 rejected records with full gate-flag detail. The same records appear in rejected.jsonl; this file adds the summary.
resolved_config.yaml	The fully-resolved run configuration (thresholds, per-domain placeholder profiles, masking profile); its SHA-256 is stamped in run_manifest.json and published as config_hash.txt.
run_manifest.json	Full generation configuration, provenance (git commit, environment), per-domain distribution, rejection counts, and the reconciliation block.

Accepted-record schema (output.jsonl)

Field names and types are taken directly from the shipped records.

Field	Type	Description
id	string (UUID)	Stable per-record identifier.
text	string	The completed document with synthetic personal-data values present at the recorded span offsets. Character offsets in spans index into this string.
category	string	The domain — one of privacy, gdpr, medical, education, transport, ecommerce, legal.
labeling	string	construction_true — labels produced by injection, not post-hoc tagging.
spans	list[object]	The ground-truth labels (see below). May be empty for a record that carries no injectable domain-allowed identifier.
masked_text	string	The anonymized view of the same document, with each injected value replaced by its <TYPE> token. This is the redacted form a de-id system should produce.

Span object:

Field	Type	Description
start / end	int	Character offsets into text (half-open). Correct by construction — the injector wrote them.
type	string	Identifier type (17 in this run — see distribution below).
value	string	The injected synthetic value present in text at [start:end].
source	string	injected — the construction-true provenance of the span.
valid	bool	Whether the injected value is a well-formed exemplar of its type. See the note below.

The two views: text vs. masked_text

For a domain-allowed identifier type, the injector writes a concrete synthetic value into text and the corresponding <TYPE> token into masked_text:

text:         … New Specialist <zfck@mail.test> …
masked_text:  … New Specialist <<EMAIL>> …
span:         {"start": 168, "end": 182, "type": "EMAIL", "value": "zfck@mail.test", "source": "injected", "valid": true}

A type the model emits that is not allowed for the record's domain is passed through as a raw <TOKEN> in both text and masked_text, and is left unlabeled by design — it is not a construction-true span. The presence of such raw tokens in masked_text is the anonymized view working as intended, not a defect: masked_text is defined as the token-bearing redacted form.

Note on valid: because the span's start/end/type come from the injector, the label is correct regardless of whether the value is well-formed — a shaped-but-invalid value is still a correctly-located, correctly-typed span. The dataset deliberately contains a mix of well-formed and shaped-but-invalid values so a detector trained on it learns to key on shape and position, not strict format validity. The valid flag exposes that distinction per span; the aggregate tally is in the evaluation report (2,246 valid / 206 not).

Rejected-record schema (rejected.jsonl)

Records are rejected at the generation gate, before de-identification injection, so they are logged in generation form rather than as {text, spans, masked_text}: id, instruction, input, output (raw, un-delivered), domain, cosine_similarity, placeholder_density, the boolean gate flags — is_hallucinated, template_leak, privacy_violation, bias_violation, gdpr_violation, url_violation, fidelity_violation — plus pii_tags and forbidden_found.

Example (representative accepted record, abridged)

{
  "id": "95fe1b66-d125-4104-8ab8-770bc15b3dd4",
  "text": "… New Specialist <zfck@mail.test>\n\n**Organization:** …",
  "category": "privacy",
  "labeling": "construction_true",
  "spans": [
    {"start": 168, "end": 182, "type": "EMAIL", "value": "zfck@mail.test", "source": "injected", "valid": true}
  ],
  "masked_text": "… New Specialist <<EMAIL>>\n\n**Organization:** …"
}

Domain distribution (accepted records)

Domain	Records
privacy	169
gdpr	168
medical	167
education	146
transport	143
ecommerce	132
legal	128
Total	1,053

Ground-truth span distribution

All 2,452 spans in output.jsonl (mean ≈ 2.3 per record). Counts match the per-type support in evaluation_report.json exactly.

Type	Spans		Type	Spans
EMAIL	585		CREDIT_CARD	57
ORDER_ID	493		ACCOUNT_NO	41
IP_ADDRESS	289		STUDENT_NAME	23
NATIONAL_ID	222		IBAN	19
CPT_CODE	190		ZIP	13
URL	161		SCHOOL_ID	13
ICD10_CODE	125		SWIFT_BIC	3
MAC_ADDRESS	85		Total	2,452
LICENSE_PLATE	72
DATE	61

---

The Veritruct Methodology

What separates this dataset from a raw generation dump is the production layer between the model and the file. Each component is documented in depth in the whitepaper; summarized here.

1. Construction-true labeling via value-packet generation + token injection

Making construction-true labeling work at generation time required solving a defect that is invisible until the labels are the product: a general-purpose model, asked for a completed record, tends to emit a skeleton — a template of fill-in-the-blank stubs — rather than a finished document. Veritruct resolves this with a division of labor:

• Personal-data fields are handled by token injection. The model emits typed placeholder tokens (<EMAIL>, <ORDER_ID>, …) at natural positions; the injector then replaces each domain-allowed token with a concrete, format-shaped synthetic value drawn from a provider library, recording the exact character span of each replacement. The recorded spans are co-referential to the injected text, so the ground-truth labels are exact by construction.
• Non-PII narrative fields are handled by a value packet — concrete non-PII values pre-generated in code and handed to the model with an instruction to use them verbatim, removing the model's excuse to emit a blank.

Together these produce a realistic completed record whose every personal-data span is correct by construction.

2. The template-leak gate — rejecting skeletons a naive gate accepts

A narrow gate that recognizes only one skeleton style silently accepts the rest and reports a falsely high yield. Veritruct's gate (in veritruct_core.generation.orchestrator) catches the full range of skeleton styles — the angle-bracket family (<INSERT_URL>, <..._HERE>), the keyword-square family ([Insert …], [TBD], [Redacted]), the slash-option family ([Sharp/Dull/Aching]), the field-label family ([Admission Date]), and bare clinical-field abbreviations ([DOB], [MRN]) — while sparing legitimate typed <UPPER_SNAKE> placeholders (which the injector depends on) and bracketed citations. On this run the template-leak gate accounted for 67 of the 197 rejections.

3. The hallucination gate — a dual-signal decision, not a cosine floor

A record is rejected as hallucinated only when two signals fail simultaneously: (1) embedding cosine similarity (via Snowflake Arctic Embed L v2.0, run 4-bit locally) below the domain threshold, and (2) a keyword-overlap safety net. A low cosine alone does not reject a record. The committed floor is 0.45 uniform across domains — a recalibration from an earlier 0.55/0.58 regime, because long completed records diverge more from their short instructions than sparse templates do and so score lower cosine-to-instruction; 0.45 recovers those legitimate completed records while still rejecting genuinely off-topic output. On this run the hallucination gate is the dominant rejection path (127 of 197).

4. Compliance masking and the hybrid detector

Masking runs against an active regulatory profile (gdpr) with an always-mask floor and severity-tiered bias/GDPR keyword filters, and the injector is domain-gated: only types a given domain permits are injected and labeled; forbidden types are passed through unlabeled. Evaluation then scores a hybrid detector — a hand-written compiled-regex registry combined with a learned GLiNER2 tier (fastino/gliner2-large-v1, confidence 0.5) — against the construction-true labels.

---

Detector Evaluation — run_cloud_20260706_000447

Because the labels are correct by construction, this evaluation grades the detector, not the labels: a low recall on a type means the detector missed spans known to be present; a low precision means it over-predicted, not that the labels are uncertain. A hybrid regex + GLiNER2 detector was scored against all 2,452 construction-true ground-truth spans across 17 identifier types (from evaluation_report.json).

Headline: micro-averaged F1 0.905 (precision 0.836, recall 0.986). Macro-averaged F1 0.811 (precision 0.828, recall 0.890).

Type	Support	Precision	Recall	F1
CPT_CODE	190	1.000	1.000	1.000
DATE	61	1.000	1.000	1.000
IBAN	19	1.000	1.000	1.000
IP_ADDRESS	289	1.000	1.000	1.000
LICENSE_PLATE	72	1.000	1.000	1.000
MAC_ADDRESS	85	1.000	1.000	1.000
ORDER_ID	493	1.000	1.000	1.000
EMAIL	585	0.998	1.000	0.999
URL	161	1.000	0.994	0.997
ICD10_CODE	125	1.000	0.952	0.975
CREDIT_CARD	57	0.950	1.000	0.974
NATIONAL_ID	222	0.937	1.000	0.967
ZIP	13	1.000	0.692	0.818
ACCOUNT_NO	41	1.000	0.537	0.698
SCHOOL_ID	13	0.140	1.000	0.245
STUDENT_NAME	23	0.055	0.957	0.104
SWIFT_BIC	3	0.000	0.000	0.000

Twelve of the seventeen types clear F1 0.96, including every high-support structured identifier. Two distinct, honestly-reported weaknesses account for the rest:

• Free-text types — high recall, low precision. STUDENT_NAME (F1 0.104; recall 0.957 but 377 false positives) and SCHOOL_ID (F1 0.245; 80 false positives) are the free-text entities the learned GLiNER2 tier is uniquely responsible for. The detector finds the entities but over-predicts their boundaries and instances. This is a genuine weakness in the free-text tier, not an averaging artifact — and because this all-domain run populates these types (which the medical-only run left near-empty), it is now measured and actionable. These two types' false positives are also the main reason micro-precision sits at 0.836 while recall is 0.986.
• Low-support recall gaps. SWIFT_BIC (support 3, all missed → F1 0.000), ACCOUNT_NO (recall 0.537), and ZIP (recall 0.692) miss a share of their spans on small supports.

Micro vs. macro. Micro pools every span (weighting each of the 2,452 equally) → F1 0.905, reflecting performance on the data that exists. Macro weights each type equally regardless of support → the low-support/free-text types pull it to 0.811. Both are reported so the reader sees the aggregate and the per-type reality.

Injected-value validity: of the 2,452 injected values, 2,246 were well-formed for their type and 206 were not (91.6% valid) — materially better than a code-heavy medical-only run. This measures the format fidelity of the injected values, not label correctness — the span offsets and types are correct by construction regardless (see the valid note above). Discussed in Limitations.

---

Yield & Rejection Economics

From run_manifest.json and rejection_audit.json:

• Overall yield: 84.2% (1,053 accepted / 1,250 attempted), reconciliation ok, across all seven domains (10 completed batches).

The 197 rejections, by gate. Gates can co-occur on a single record, so the per-gate counts sum to more than the 197 unique rejected records:

Gate	Count
Hallucination	127
Template-leak (skeleton output)	67
Bias	8
Fidelity (placeholder)	7
GDPR (severe)	5
Privacy	0
URL	0
Unique rejected records	197

Hallucination and template-leak together account for the overwhelming majority of rejections — the two gates that do the most quality work on completed de-identification records.

---

Provenance & Reproducibility

Generation (Veritruct Cloud substrate)

Component	Specification
Modal app / serving class	syntheval-cloud-vllm / MistralNemoAWQVllm
GPU	NVIDIA A10G (24 GB)
Serving engine	vLLM 0.8.5.post1 (remote Modal container: PyTorch 2.6.0, Transformers 4.51.3)
Generation model	Mistral-Nemo-Instruct-2407 AWQ (in-house; Mistral-Nemo-Instruct-2407-AWQ_20260529_152601)
max_model_len / RPC timeout	2,560 / 600 s
Sampling	temperature 0.3 · top_p 0.92 · repetition_penalty 1.1 · max_new_tokens 1,024
Batch size / workers / seed	128 / 1 / 99
Over-generation ratio	0.25
Target prompts	1,000 (1,053 accepted after surplus-retention; valid surplus records are kept, not trimmed)

Gating, masking & evaluation (local)

Component	Specification
Embedding model	Snowflake Arctic Embed L v2.0, 4-bit (in-house) — runs locally even though generation is remote, so the quality signal stays on operator-controlled hardware
Hallucination thresholds	0.45 uniform across all domains (config-declared)
Masking profile	gdpr · substitute PII: true
Detector	hybrid — hand-written compiled-regex registry + GLiNER2 (fastino/gliner2-large-v1), confidence 0.5
Labeling mode	construction_true

Run identity & code provenance

• Run ID: run_cloud_20260706_000447
• Product: deid · Labeling: construction_true · Mode: production
• Generated at: 2026-07-06T00:27:18
• Git: commit 19bc8bfd1acc6c6a32039f19e5eba9913763b69b, branch main, clean tree (committed before generation)
• Resolved config: resolved_config.yaml, SHA-256 23f23eb0…4bf74 (published as config_hash.txt)
• Reconciliation: ok — accepted count, rejected count, and the seven-domain distribution all agree (1,053 across 10 completed batches).

This is the all-domain Cloud de-identification run documented in the whitepaper (§8.5).

---

Intended Uses

• Training and benchmarking de-identification / PII-redaction models — NER-style span taggers that must locate and type personal data across regulated-domain text, with construction-true labels usable as ground truth.
• Evaluating existing detectors across a broad identifier vocabulary — 17 identifier types spanning seven domains; because the labels are correct by construction, detector scores grade the detector, not the labels.
• Compliance- and privacy-aware NLP research — the dataset carries synthetic identifiers only; the masked_text view supplies the redaction target directly.

Limitations (stated as plainly as the results)

Mirroring the whitepaper's limitations section:

1. Free-text detector precision. The learned tier recovers STUDENT_NAME and SCHOOL_ID with high recall but low precision (over-prediction) — a genuine, measured weakness on this run, not an averaging artifact.
2. Low-support types. SWIFT_BIC, ACCOUNT_NO, and ZIP carry small supports here and show recall gaps; their detector scores are less stable than the high-support types.
3. Injected-value format fidelity. 91.6% of injected values are well-formed; span labels are correct by construction regardless, but a use case needing strictly valid exemplars should filter on the valid flag.
4. Pattern-plus-learned detection cannot be proven complete. Detector metrics characterize the failure modes exercised by this corpus, not an exhaustive guarantee.
5. Single generation model / language. English, generated by one 12B model (AWQ-quantized); characteristics of that model are present throughout.
6. Synthetic content is illustrative. All names, dates, codes, and identifiers are fabricated; nothing refers to real individuals, records, or encounters.

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0) — a PBH Veritruct Showcase release.

This dataset is free to use, modify, and redistribute — including commercially — under one condition: attribution to PBH Applied Systems, LLC. If you train on it, fine-tune with it, benchmark against it, or build a derivative dataset from it, credit PBH Applied Systems and link back to this release (the citation below satisfies this). Attribution is the only string attached, and it is deliberate: this is a showcase release, and seeing where the data travels is the point.

A few facts relevant to licensing, stated plainly so the terms are transparent rather than buried:

• The records are synthetic, generated by Mistral-Nemo-Instruct-2407 (Apache 2.0); no web-scraped or copyrighted source text was used as input, and the dataset contains no real personal or health information by design.
• CC BY 4.0 covers the data in this repository. The Veritruct methodology, fixtures, masking profiles, gating thresholds, value-injection providers, and the Veritruct Studio/Cloud pipelines are proprietary to PBH Applied Systems, LLC and are not included in, nor licensed by, this release. You may use the data freely; you do not receive a license to the system that produced it.
• This is the publicly licensed showcase sample. Bespoke, larger-scale, or domain-tailored datasets produced under the same methodology are delivered under separate commercial engagement terms — see below.

The Hub renders this as the recognized CC BY 4.0 badge (with filtering and the standard legal text). The "PBH Veritruct Showcase" framing is branding carried in this section, not a custom metadata identifier — the recognized license keyword and a custom license_name cannot coexist in the card metadata.

---

Citation

@misc{hill2026veritruct,
  title        = {Veritruct: Quality-Gated Synthetic Data Generation for Regulated Industries},
  author       = {Hill, Patrick},
  year         = {2026},
  howpublished = {PBH Applied Systems, LLC},
  note         = {Dataset release: run\_cloud\_20260706\_000447 (construction-true regulated-domain de-identification)},
  url          = {https://pbhappliedsystems.com}
}

---

About PBH Applied Systems

PBH Applied Systems, LLC is an Oklahoma City–based applied machine learning and AI systems company specializing in quality-gated synthetic data generation, production-grade model evaluation, quantization pipelines, and agentic AI infrastructure — with an emphasis on engineering rigor, reproducibility, and real-world deployment constraints.

Founder — Patrick Hill, M.S. — Principal AI/ML Systems Architect; M.S. in Software Engineering (AI/ML concentrations); author of Applied Machine Learning: Concepts, Tools, and Case Studies.

Need defensible synthetic data?

Veritruct produces datasets that ship with the verification a regulated buyer would otherwise have to perform themselves: construction-true labels, a detector evaluation scored against known-correct spans, rejection-reason breakdowns, an injected-value validity tally, and a reconciliation of the run's own counts.

👉 Discuss a synthetic-data engagement — volume, domains, compliance frameworks, and delivery timeline.

🌐 Website	pbhappliedsystems.com
🤖 Live AI Agent Demo	pbhappliedsystems.com/assistant.html
📧 Email	patrick@pbhappliedsystems.com
💼 LinkedIn	PBH Applied Systems, LLC
▶️ YouTube	@pbhappliedsystems

---

Generated, gated, masked, labeled, and evaluated by PBH Applied Systems, LLC · Veritruct Cloud · Run run_cloud_20260706_000447 · Every figure on this card is a field in the artifacts shipped with the data.