beyond369/Privasis-USA-Curated-37k

Privasis-Technical-Curated-37k

A curated subset of NVIDIA's nvidia/Privasis-USA (specifically the corpus split), engineered for domain-specific pre-training across five technical/enterprise pillars.

Dataset Summary

• Total Documents: 37,000
• Estimated Token Volume: ~48.63 Million tokens ($\text{characters} / 4$)
• Document Length: Bounded between 500 and 16,000 characters (~125 to ~4,000 tokens)
  • Mean Tokens / Doc: 1,314.3
  • Median Tokens / Doc: 1,221.0
• Persona Diversity: 100% unique (37,000 unique personas across 37,000 documents; zero demographic repetition).

---

Domain & Tag Distribution

The dataset strictly balances five high-density technical and operational domains:

Primary Bucket	Target Share	Document Count	Est. Tokens	Primary Focus
research	25%	9,250	~14.5M	Experimental methodologies, telemetry, technical proposals, sensor stacks
tech	25%	9,250	~13.0M	Systems architecture, hardware fabrication, incident logs, SOPs, DevOps tickets
legal	20%	7,400	~9.6M	Regulatory compliance, affidavits, statutory forms (IRS Form 990), insurance policies
medical	15%	5,550	~5.7M	Clinical notes, patient correspondence, orthopedic evaluations, rehabilitation plans
finance	15%	5,550	~5.8M	Balance sheets, accounting logs, budgets, transaction ledgers

---

Curation Methodology

This dataset was derived using multi-tag probabilistic downweighting and persona deduplication:

1. Length Bounding: Documents with $< 500$ characters (stubs, broken rows) or $> 16,000$ characters (extreme raw dumps, runaway CSVs) were dropped to align with standard context windows (2k/4k tokens).
2. Anti-Contamination Downweighting: Rather than hard-blocking co-occurring tags, secondary tags were penalized multiplicatively during reservoir sampling:
   • Severe downweighting ($0.05\times$ acceptance odds): notes, marketing, sales, other
   • Moderate downweighting ($0.50\times$ acceptance odds): admin, project, hr, creative, educational This cleanly purged bake-sale flyers, low-effort youth summaries, and repetitive committee minutes while retaining authentic operational and technical records.
3. 1:1 Persona Deduplication: Deduplicated on nemotron_personas_usa_uuid so that every synthetic individual appears at most once in the entire corpus, preventing identity memorization.

---

Note on Quality Classifiers & Pre-training Gaps

FineWeb-Edu Score Divergence

250 records from this curated subset were empirically evaluated using HuggingFaceFW/fineweb-edu-classifier ($N=250$ representative sample, 0–5 scale):

• Mean Score: 1.02 / 5.0 (Median: 0.98, Std: 0.40)
• Score Range: 0.11 to 2.68
• Records Scoring Over 3 (Standard FineWeb cutoff): 0.0%

Domain Pillar	Sample Count	Mean Score	Min / Max
Research	72	1.25	0.54 / 2.68
Tech	61	1.13	0.34 / 2.28
Finance	39	0.89	0.11 / 1.56
Medical	33	0.83	0.35 / 1.44
Legal	45	0.73	0.15 / 1.52

Filling an Enterprise Pre-training Gap

The FineWeb-Edu classifier is calibrated specifically for didactic textbook prose (e.g., historical essays, Wikipedia articles, introductory STEM tutorials). Standard pre-training filtering pipelines that enforce an educational score cutoff (such as $\ge 3.0$) discard 100% of these records. This subset deliberately rescues that dense, non-conversational operational data—providing critical exposure to statutory legal phrasing, hardware debugging telemetry, dosage titrations, and business schemas.

---

Intended Use

Designed as a high-density, multi-domain foundational supplement. When training general-purpose or enterprise LLMs, blending this subset at 2% to 5% of a continued pre-training mix may help models ground reasoning over structured corporate documentation, clinical logs, legal schedules, and hardware specifications.

---

Source & Attribution

Derived from NVIDIA's Privasis-USA dataset under the CC-BY-4.0 license:

• Original Dataset: nvidia/Privasis-USA
• Paper: "Privasis: Synthesizing the Largest 'Public' Private Dataset from Scratch" (NVIDIA Research)