pbhappliedsystems/veritruct-studio-regulated-instruct-1K

Veritruct Studio — Regulated-Domain Synthetic Instruction Dataset

Run run_studio_20260705_203339 · Flagship release demo dataset · PBH Applied Systems, LLC

Generated, gated, masked, 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 Attested Synthetic Data Generation.

📋 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. This card is the deliberate inverse. Every record in this dataset cleared a documented Veritruct quality-and-compliance cascade — semantic hallucination filtering, a template-leak gate, regulation-specific PII masking, automated evaluation, and a hybrid leak audit — 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 that produced these records is documented in full in the whitepaper, Veritruct: Quality-Gated Synthetic Data Generation for Regulated Industries (Hill, 2026).

---

TL;DR

Records (accepted, delivered)	1,014
Records rejected (logged, not delivered)	108
Total attempted	1,122
Overall yield	90.4%
Domains	7 — ecommerce, privacy, medical, transport, legal, education, GDPR
Format	Instruction / input-context / masked-output JSONL
Residual PII leak rate (hybrid detector)	0.0%
Placeholder coverage	100%
Lexical diversity (MATTR)	0.790
Mean accepted instruction–output cosine	0.57
Near-duplicate rate	0.0%
Substrate	Veritruct Studio — local NVIDIA RTX 4090 (in-process)
Generation model	Mistral-Nemo-Instruct-2407, 4-bit BitsAndBytes (in-house)
Reconciliation status	ok — all counts and the domain distribution agree

---

What This Dataset Is

This is a synthetic instruction-following dataset for regulated, compliance-heavy domains. Each record is a professional-context instruction paired with a generated long-form response — a GDPR processor-onboarding checklist, a HIPAA breach-investigation report, an ALPR procurement-policy recommendation, a medical-billing audit, a data-subject-rights fulfillment record, and so on — produced from the perspective of a named professional persona (Compliance Officer, DSAR Processing Manager, Cross-Border Transfer Specialist, Learning Analytics Researcher, ALPR System Engineer, Financial Crime Investigator, and others).

Every delivered record has:

• cleared a dual-signal hallucination gate (the output is on-topic with respect to its instruction),
• cleared a template-leak gate (the output is a completed record, not a fill-in-the-blank skeleton),
• had every detected piece of personal data replaced with a typed placeholder (<EMAIL>, <NATIONAL_ID>, <CPT_CODE>, <IP_ADDRESS>, <LICENSE_PLATE>, …) in place, and
• survived a post-run leak audit — a hybrid of a hand-written compiled-regex registry and a learned GLiNER2 NER tier — that scans the placeholder-stripped remainder of every record for raw personal data the masker might have missed.

The records that did not survive are not silently dropped — they are written to rejected.jsonl with the gate flags that caused rejection, and summarized in rejection_audit.json. The clean delivered dataset is the product of removal at the gate, not of constraining the generator.

This release was generated on Veritruct Studio, the local-GPU substrate of the Veritruct methodology — generation runs in-process on a single consumer RTX 4090 at near-zero marginal cost. Its sibling, Veritruct Cloud, offloads generation to a Modal A10G served by vLLM while importing the identical shared core: gates, masking, evaluation, and audit are the same code on both substrates. A single whitepaper documents the whole system; the substrates differ only in where generation tokens are produced.

---

Dataset Structure

Files

File	Contents
README.md	This dataset card.
config_hash.txt	The SHA-256 checksum of resolved_config.yaml (36e83426…16ef9) — 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 evaluation report (diversity, quality, compliance, yield, clustering).
output.jsonl	The 1,014 accepted, delivered records. One JSON object per line.
rejected.jsonl	The 108 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 108 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, 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.
instruction	string	The task, framed from a named professional persona's perspective.
input	string	Structured context: Domain: … | Framework/Jurisdiction: … | Persona: … | Focus: ….
output	string	The generated long-form response, after PII masking (placeholders substituted in place).
category	string	The domain: one of ecommerce, privacy, medical, transport, legal, education, gdpr.
output_format	string	The response template family (e.g. token_rich, patient_intake, vendor_marketplace, content_and_seo, freight_and_logistics).
cosine_similarity	float	The instruction–output embedding cosine, computed on the raw (pre-masking) output. This is the hallucination-gate quality signal, persisted per record.
placeholder_density	int	Count of typed placeholders in the masked output.
placeholder_types	list[string]	The distinct placeholder types present (e.g. ["ICD10_CODE","CPT_CODE","DATE","NATIONAL_ID"]).

Note on cosine_similarity: the score is computed on the model's raw output, before masking substitutes any spans, so it reflects the actual generated response rather than its post-masking form. This is why a record can carry both a meaningful semantic score and a fully masked body.

Rejected-record schema (rejected.jsonl)

Rejected records carry per-gate flags so the rejection is fully auditable: id, instruction, input, output (raw, un-delivered), domain, cosine_similarity, placeholder_density, and the boolean gate flags — is_hallucinated, template_leak, privacy_violation, bias_violation, gdpr_violation, url_violation, fidelity_violation — plus pii_tags and forbidden_found. Note the rejected schema keys the domain as domain (the accepted schema uses category).

Example (representative accepted record, abridged)

{
  "id": "051dd815-ecd3-4719-87ef-b3d70d93af1c",
  "instruction": "As Financial Crime Investigator at Fintech compliance unit would frame it: Write an SEO content brief on payment facilitator sub-merchant onboarding and KYB requirements, including search intent, product-page metadata, internal linking targets, <URL> structure, and content quality criteria.",
  "input": "Domain: E-Commerce & Payments | Framework: PCI-DSS/GDPR | Persona: Financial Crime Investigator | Focus: payment facilitator sub-merchant onboarding and KYB requirements",
  "output": "… \"iban\": \"<IBAN>\", \"email\": \"<EMAIL>\", \"ip_address\": \"<IP_ADDRESS>\" …",
  "category": "ecommerce",
  "output_format": "content_and_seo",
  "cosine_similarity": 0.845215,
  "placeholder_density": 6,
  "placeholder_types": ["IBAN", "EMAIL", "IP_ADDRESS"]
}

Domain distribution (accepted records)

Domain	Records	Per-domain yield
medical	169	93.4%
privacy	160	93.6%
education	142	88.8%
gdpr	142	87.7%
legal	140	89.7%
transport	136	88.3%
ecommerce	125	90.6%
Total	1,014	90.4%

Domains are intentionally not perfectly balanced in the delivered file. Valid records that pass every gate are never discarded to force balance; domain rebalancing is a downstream post-processing step. Generation targeted near-equal domain weights (0.14–0.15 each); the delivered counts reflect per-domain yield differences.

---

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

The gate rejects a record as hallucinated only when two signals fail simultaneously:

1. Embedding cosine similarity. A vector is computed for the instruction and for the raw output using Snowflake Arctic Embed L v2.0 (run 4-bit, locally), and the cosine between them is taken. This is the first signal; it fails when it falls below the domain threshold. The cosine is computed on the raw output, before masking, so it reflects the model's actual response.
2. Keyword-overlap safety net. A conservative check that the output shares substantive tokens with the instruction — confirming the output is at minimum anchored to the instruction's subject.

A record is rejected as hallucinated only if (cosine check fails) AND (keyword-overlap check fails).

A low cosine alone does not reject a record. This run demonstrates the rule directly: the minimum accepted cosine is 0.34 — below the 0.45 floor — retained because its keyword-overlap signal passed. The cosine threshold is a necessary but not sufficient condition for rejection.

The 0.45 unified floor. The committed Studio configuration sets the cosine floor at 0.45 for the default and every per-domain threshold (uniform). This is a recalibration from an earlier 0.55/0.58 regime: as generation quality shifted toward long, fully-completed records, those richer outputs scored lower cosine-to-instruction than sparse templates did (a long completed record diverges more from its short instruction than a stub does). Under the old floor, legitimate completed records in the 0.45–0.58 band were being false-rejected as hallucinations; lowering the floor to 0.45 recovers them while still rejecting genuinely off-topic output, which scores below 0.45 and fails both signals. Thresholds are declared in configuration, not hard-coded.

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

A general-purpose model, asked for a completed record, frequently emits a skeleton: a template studded with fill-in-the-blank stubs ([Your Name], [Admission Date], <INSERT_PATIENT_IDS_HERE>) where concrete content should be. 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) uses two compiled patterns to catch 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 and bracketed citations (guarded by a four-digit-year / "et al." exclusion). In this run the template-leak gate was the single largest rejection category (63 of 108).

3. The compliance architecture — a gate, not a disclaimer

• Layered PII masking (pii_masker.py) against an active regulatory profile (gdpr for this run): detected personal data is replaced with typed placeholders in place. An always-mask floor covers phone numbers, email addresses, SSNs, credit-card numbers, and IP addresses in every domain regardless of that domain's allow-list — visible in this run's placeholder counts, where PHONE (1,463) and SSN (193) appear despite not being domain-preferred types.
• Severity-tiered keyword filters — a bias filter (severe tier rejects everywhere; moderate tier is exempted where the terminology is the legitimate subject) and a GDPR keyword filter (severe rejects; moderate is informational and surfaces as the GDPR-surface metric).
• Placeholder fidelity — rejects records carrying a forbidden placeholder type for their domain, and routes under-floor-density records toward regeneration rather than acceptance.

4. Hybrid leak auditing — the verification does not trust its own masker

The leak audit combines a hand-written compiled-regex pattern registry with a learned GLiNER2 NER tier (fastino/gliner2-large-v1, confidence 0.5) and scans the placeholder-stripped remainder of every delivered record. Using detection independent of the masker's own patterns is deliberate: a detector that reuses the masker's patterns can report clean while a value the masker's patterns missed is still present. This run's hybrid detector found a residual PII leak rate of 0.0% across all 1,014 delivered records.

5. The attested evaluation report — a deliverable, not a diagnostic

The evaluation pipeline runs automatically after generation, evaluates the entire accepted corpus (it does not sample), reconciles the counts, and ships a machine-readable evaluation_report.json. It is what makes every quality claim on this card checkable by a third party rather than asserted.

---

Evaluation Report — run_studio_20260705_203339

All figures below are read directly from evaluation_report.json (full evaluation completed in 11.64 s; reconciliation status ok; not degraded).

Yield & rejection economics

• Overall yield: 90.4% (1,014 accepted / 1,122 attempted)
• Hallucination rate (vs. total attempts): 4.01%
• Accepted cosine: mean 0.568, median 0.549, min 0.342, max 0.968, std 0.113

The 108 rejections, by gate (from run_manifest.json). Gates can co-occur on a single record, so the per-gate counts sum to more than the 108 unique rejected records:

Gate	Count
Template-leak (skeleton output)	63
Hallucination	45
Fidelity (placeholder)	5
GDPR (severe)	3
Bias (severe/moderate)	2
Privacy	0
URL	0
Unique rejected records	108

Diversity

Metric	Value	Note
MATTR (100-token window)	0.790	Scale-invariant headline diversity grade
Vocabulary size	7,009	Unique types
Hapax ratio	0.308	Fraction of vocabulary appearing exactly once
Top-20 token concentration	0.113	Low concentration → no small-vocabulary dominance
Near-duplicate rate	0.0%	Over instructions
Bigram / trigram ratio	0.604 / 0.822	Reported as size-sensitive density, not the diversity grade
Unigram ratio	0.101	Size-sensitive; falls mechanically with corpus size — not a quality signal

The raw type-token ratios are explicitly relabeled as size-sensitive density rather than presented as diversity. MATTR, hapax ratio, and top-20 concentration are the scale-robust grades.

Quality (well-formedness)

Metric	Value
Output length (tokens) — mean / median	496 / 492
Output length — p5 / p95	341 / 669
Output length — min / max	29 / 954
Char length — mean	2,745
Token-length entropy (Shannon)	3.08
Sentence-level repetition rate	1.6%
Short-output rate	0.1%
Instruction–output cosine — mean / median	0.57 / 0.55

Compliance (measured on the delivered set)

Metric	Value
Residual PII leak rate (hybrid detector)	0.0%
Echo-contamination rate	0.0%
Placeholder coverage	100%
Placeholder density (mean)	5.24
GDPR-surface rate	35.6%
Bias-flag rate	0.3%

Most frequent placeholder types: EMAIL (1,525), PHONE (1,463), ORDER_ID (456), CPT_CODE (329), IP_ADDRESS (276), NATIONAL_ID (236), SSN (193), URL (181), ICD10_CODE (148), DATE (109), IBAN (77), ZIP (71), MAC_ADDRESS (54), LICENSE_PLATE (53), CREDIT_CARD (24).

Reading coverage/density: every delivered record carries at least one placeholder (100% coverage) at a mean density of 5.24. The PHONE and SSN counts — types no domain lists as preferred — are the always-mask floor doing its job. The leak rate (0.0%) is the metric that proves no raw personal data survived.

Clustering (corpus-level integrity)

Unsupervised k-means over record embeddings auto-selected k = 6. Per-cluster domain purity ranges 0.42–1.00, confirming most clusters map cleanly onto single domains:

Cluster	Dominant domain	Purity	Top terms
3	education	1.00	student, students, learning, privacy
2	ecommerce	0.99	fraud, transaction, iban, transactions
0	legal	0.97	legal, compliance, case, review
5	medical	0.94	patient, date, code, compliance
4	transport	0.91	alpr, license, plate, traffic
1	privacy + gdpr	0.42	data, processing, access, compliance

(Silhouette 0.073, inertia 496.06, PCA variance explained 0.109 — k-means over record embeddings. The privacy/gdpr blend in cluster 1 is expected: the two domains share regulatory vocabulary by construction.)

---

Provenance & Reproducibility

Generation (Veritruct Studio substrate)

Component	Specification
Backend	LocalHFBackend — in-process generation (no remote serving)
GPU	NVIDIA RTX 4090 (24 GB)
Host environment	WSL2 · Python 3.10.12
Framework	PyTorch 2.4.1 + cu121 · Transformers 4.57.6
Generation model	Mistral-Nemo-Instruct-2407, 4-bit BitsAndBytes (in-house quantized)
Sampling	temperature 0.3 · top_p 0.92 · repetition_penalty 1.1 · max_new_tokens 1,024
Batch size / workers / seed	22 / 1 / 99
Over-generation ratio	0.25
Target prompts	1,000 (1,014 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) — run simultaneously with the generation model on the same RTX 4090
Hallucination thresholds	0.45 uniform across all domains (config-declared)
PII detection & masking	Hand-written compiled-regex pattern registry (Python re)
Learned NER tier	GLiNER2 (fastino/gliner2-large-v1), confidence 0.5, include_spans=True — the hybrid detector
Masking profile	gdpr · substitute PII: true

Run identity & code provenance

• Run ID: run_studio_20260705_203339
• Product: instruction · Mode: production
• Generated at: 2026-07-05T22:50:56
• Git: commit 19bc8bfd1acc6c6a32039f19e5eba9913763b69b, branch main, clean tree (committed before generation)
• Resolved config: resolved_config.yaml, SHA-256 36e83426…16ef9
• Reconciliation: ok — output.jsonl line count, rejected.jsonl line count, and run_manifest.json totals/domain distribution all agree.

This is the Veritruct Studio instruction flagship run. The shared methodology is validated at ten-thousand-record scale in the whitepaper on per-substrate medical de-identification runs (86.9% Studio, 85.5% Cloud), with a hybrid detector micro-averaged F1 of 0.964 against 49,344 construction-true ground-truth spans.

---

Intended Uses

• Instruction-tuning and SFT for assistants operating in regulated, compliance-heavy professional contexts (legal, healthcare, data protection, surveillance/transport oversight, EdTech, e-commerce compliance).
• Compliance- and privacy-aware generation research — the typed-placeholder masking makes the dataset usable where raw PII would be a non-starter.
• A reference example of an attested synthetic-data deliverable — the data plus its machine-readable evaluation and provenance artifacts.

Limitations (stated as plainly as the results)

These are lifted from the whitepaper's limitations section and apply to this dataset:

1. Pattern-plus-learned compliance cannot be proven complete. The 0.0% residual leak rate is a measured result from a hybrid regex + GLiNER2 detector, not a guarantee of zero PII. It is reported as measured and audited, never as proven perfect.
2. Thresholds are calibrated, not self-adjusting. The 0.45 hallucination floor was empirically calibrated to this pipeline's embedding distribution; it is not adaptive.
3. The bias-flag metric is a reported residual. The 0.3% bias-flag rate reflects legitimate professional terminology surfaced for reporting rather than rejection — not suppressed.
4. Demonstrated at the thousand-record scale. This is a 1,014-record instruction run; the methodology is separately validated at 10K scale on de-identification runs in the whitepaper.
5. Single generation model / language. Outputs are English, generated by one 12B model (4-bit quantized); characteristics of that model are present throughout.
6. Synthetic content is illustrative. Names, dates, codes, and identifiers in the outputs are fabricated or masked; nothing in the data refers to real individuals, cases, or records.

---

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, so nothing upstream restricts redistribution.
• All detected personal data is masked to typed placeholders; the dataset contains no real PII by design.
• CC BY 4.0 covers the data in this repository. The Veritruct methodology, fixtures, masking profiles, gating thresholds, 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\_studio\_20260705\_203339},
  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: rejection-reason breakdowns, scale-invariant diversity, placeholder coverage, an audited leak rate, 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, and evaluated by PBH Applied Systems, LLC · Veritruct Studio · Run run_studio_20260705_203339 · Every figure on this card is a field in the evaluation_report.json shipped with the data.