Post
2322
🚀 New from Veritruct — and an argument about what a dataset card should be. Most synthetic datasets on the Hub ship row counts, a license, and little else. De-identification corpora are worse: their span labels are usually unverifiable. We published the opposite — two datasets, two record types, one quality-gated methodology.
① Construction-true de-identification — labels correct by construction, because every PII value is injected at an offset the pipeline records. Ground truth, not fallibly-tagged.
👉 pbhappliedsystems/veritruct-cloud-regulated-deid-1k
1,053 records · 7 domains · 2,452 ground-truth spans · 17 identifier types. A hybrid regex+GLiNER2 detector scores micro-F1 0.905 against the known-correct labels — and we report the free-text precision gaps, not just the wins.
② Regulated-domain instruction — same gates, same provenance.
👉 pbhappliedsystems/veritruct-studio-regulated-instruct-1K
1,014 records · 90.4% yield · MATTR 0.790 · 0.0% residual PII leak.
Every record cleared a documented cascade — dual-signal hallucination gate, template-leak gate, layered PII masking — and every number on each card is a field in the evaluation_report.json shipped beside the data. Rejections ship too, each tagged with its failing gate. Two substrates: Studio (local GPU) · Cloud (Modal + vLLM).
📄 Whitepaper: https://pbhappliedsystems.com/Veritruct_Quality-Gated_Synthetic_Data_Generation_for_Regulated_Industries.pdf
🔎 Overview: https://pbhappliedsystems.com/veritruct.html
CC BY 4.0 — commercial use welcome, just credit it. Need defensible synthetic data at scale? Let's talk.
— Patrick Hill, PBH Applied Systems
① Construction-true de-identification — labels correct by construction, because every PII value is injected at an offset the pipeline records. Ground truth, not fallibly-tagged.
👉 pbhappliedsystems/veritruct-cloud-regulated-deid-1k
1,053 records · 7 domains · 2,452 ground-truth spans · 17 identifier types. A hybrid regex+GLiNER2 detector scores micro-F1 0.905 against the known-correct labels — and we report the free-text precision gaps, not just the wins.
② Regulated-domain instruction — same gates, same provenance.
👉 pbhappliedsystems/veritruct-studio-regulated-instruct-1K
1,014 records · 90.4% yield · MATTR 0.790 · 0.0% residual PII leak.
Every record cleared a documented cascade — dual-signal hallucination gate, template-leak gate, layered PII masking — and every number on each card is a field in the evaluation_report.json shipped beside the data. Rejections ship too, each tagged with its failing gate. Two substrates: Studio (local GPU) · Cloud (Modal + vLLM).
📄 Whitepaper: https://pbhappliedsystems.com/Veritruct_Quality-Gated_Synthetic_Data_Generation_for_Regulated_Industries.pdf
🔎 Overview: https://pbhappliedsystems.com/veritruct.html
CC BY 4.0 — commercial use welcome, just credit it. Need defensible synthetic data at scale? Let's talk.
— Patrick Hill, PBH Applied Systems