Hugging Face's logo

Datasets:
xorushi
/
secureflow-ai-datasets

Dataset card Files
xet
 Community
1
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('text', {}), NamedSplit('validation'): ('text', {}), NamedSplit('test'): ('pdffolder', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SecureFlow AI — Comprehensive Dataset Architecture & Catalog

This document provides complete, publication-grade specifications for every dataset in the SecureFlow AI Data Leakage Prevention (DLP) ecosystem. It covers dataset origins, exact schemas, security classifications, compliance mappings, preprocessing transformations, and evaluation benchmark roles.

⚡ Quickstart Dataset Reproduction

To reproduce all datasets, benchmarks, and baseline models in a single command:

cd ai-service/data/scripts

# 1. Master one-click end-to-end reproduction (download -> process -> showcase -> train)
python reproduce_all.py --all

# Or download all raw datasets from your personal Hugging Face repository snapshot:
python download_raw_datasets.py --hf-repo <your-username>/secureflow-ai-datasets

# Or upload your local raw datasets to your Hugging Face account:
python upload_to_huggingface.py --repo-id <your-username>/secureflow-ai-datasets --private

# Or run individual stages:
# Download / verify raw datasets only
python download_raw_datasets.py --dataset all

# Build all processed splits and benchmarks
python organize_datasets.py --target all

# Run specific single-dataset targets:
python organize_datasets.py --target document_classification
python organize_datasets.py --target pii_ner
python organize_datasets.py --target um_dlp
python organize_datasets.py --target contextual
python organize_datasets.py --target enron
python organize_datasets.py --target ocr_corpus

# View structured validation showcase table
python present_datasets.py

# Train baseline 4-tier document classifier
python train_classifier.py

☁️ Uploading to Your Personal Hugging Face Dataset Repo

If you want to host all 5.12 GB of raw datasets on your own Hugging Face account (free and private):

  1. Log in to Hugging Face:
    pip install huggingface_hub
huggingface-cli login
  2. Upload your local raw folder:
    cd ai-service/data/scripts
python upload_to_huggingface.py --repo-id <your-username>/secureflow-ai-datasets --private
  3. Download on any fresh machine:
    python download_raw_datasets.py --hf-repo <your-username>/secureflow-ai-datasets
python organize_datasets.py --target all

🏛️ System Data Hierarchy 

ai-service/data/
├── raw/                                      # Cleaned, immutable raw data archives
│   ├── disc/                                 # DISC.json (35.5 MB)
│   ├── medical_phi/                          # HIPAA PHI Parquet file (1.2 MB)
│   ├── roberta_pii_synth/                    # 120k Arrow PII dataset
│   ├── um_dlp_benchmark/                     # 1,343 Arrow DLP records
│   ├── contextual_sensitive/                 # 1,000 Arrow context records
│   ├── stargate_pdfs/                        # 7,394 Scanned PDF documents
│   └── enron_emails/                         # enron_emails_raw.csv (1.7 MB)
│
├── processed/                                # CORE MODEL TRAINING PARTITIONS
│   ├── document_classification/              # 4-Tier Sensitivity Classifier
│   │   ├── train.csv                         # 5,330 rows (70% Stratified Training Set)
│   │   ├── validation.csv                    # 1,142 rows (15% Stratified Validation Set)
│   │   └── test.csv                          # 1,143 rows (15% Stratified Test Set)
│   │
│   └── pii_ner_training/                     # Transformer NER Token Classification
│       ├── train.jsonl                       # 96,000 PII training spans
│       ├── validation.jsonl                  # 12,000 PII validation spans
│       └── test.jsonl                        # 12,000 PII test spans
│
├── benchmarks/                               # INDEPENDENT EVALUATION SUITES
│   ├── dlp_robustness/
│   │   └── um_dlp_test.csv                   # 1,343 DLP adversarial & noise records
│   ├── contextual_sensitivity/
│   │   └── contextual_test.csv               # 1,000 Context ambiguity records
│   ├── corporate_generalization/
│   │   └── enron_test.csv                    # 2,000 Real corporate emails from AESLC
│   ├── pii_extraction/
│   │   └── pii_benchmark_12k.jsonl           # 12,000 Character-level PII test spans
│   └── ocr_pipeline/
│       └── pdf_test_corpus/                  # 7,394 STARGATE PDFs for OCR testing
│
└── scripts/                                  # Centralized data processing scripts
    ├── organize_datasets.py                  # Automated reorganizer, cleaner & splitter
    └── train_classifier.py                   # Baseline 4-tier model training script

📋 Comprehensive Dataset Specifications

1. DISC (Declassified Intelligence Security Corpus)

  • Local Path: ai-service/data/raw/disc/DISC.json
  • Primary Processed Target: ai-service/data/processed/document_classification/
  • Modality & Format: Single JSON file (~35.5 MB)
  • Volume: 2,459 documents
  • Origin / Provenance: Digital National Security Archive (DNSA) declassified intelligence cables and diplomatic memos (Department of State, CIA, DIA, JCS).
  • Compliance & Threat Model: Defense security clearance mapping, government classified information leakage, defense-grade access control (DoD 5200.01 / ISO 27001).

Schema Definition

Field Type Description Sample Value
DocID integer Unique record index 1
Title string Intelligence report subject title "Mujahedin Cross Border Cow Raid into Soviet Union..."
Text string Cleaned full body text "Net VOW IDE 1% vue Of oo il cause..."
OCRtext string Raw OCR extraction containing noise "Net VOW IDE 1% vue Of oo il cause ~ wes..."
Abstract string Human-curated document summary "Soviet Union Armed Forces retaliate..."
Classification list[dict] Security tags and release dates [{"ClassID": "1_1", "Label": "Top Secret", "Date": "June 14, 1987"}, {"ClassID": "1_2", "Label": "Unclassified"}]
Database string Source repository "Digital National Security Archive"
Domain string Topic domain "Afghanistan: The Making of U.S. Policy, 1973-1990"
Author string Reporting agency "United States Consulate. Peshawar"
StoreId integer Storage index 1679059219

SecureFlow AI Mapping & Processing

  • Clearance Mapping:
    • Top Secret, SCI, Restricted $\longrightarrow$ Highly Confidential
    • Secret, Confidential $\longrightarrow$ Confidential
    • Official Use Only, FOUO $\longrightarrow$ Internal
    • Unclassified $\longrightarrow$ Public
  • Watermark Stripping: Regex removes leading security headers (e.g. TOP SECRET, SECRET PESHAWAR 1084) to prevent model shortcut learning.

2. Medical PHI (Protected Health Information)

Schema Definition

Field Type Description Sample Value
prompt string Patient question detailing symptoms & history "last year my wife was went through a surgery for appendix cancer..."
completion string Physician response with clinical diagnosis & advice "Hi and welcome to HCM. First, you dont have to worry. This cant be tumour..."

SecureFlow AI Mapping & Processing

  • Classification Assignment: Mapped to Confidential (Healthcare/PHI domain).
  • Feature Extraction: Combined prompt + " " + completion into unified narrative text for the 4-tier document classifier.

3. RoBERTa-PII-Synth (Synthetic PII Token NER)

Schema Definition

Field Type Description Sample Value
text string Raw noisy sentence containing synthetic PII "DON'T SHACE buX uZrE's 1970rodney.lewis'S coMatctD sarahperez@aol.com /g2118x174 / ssn 0651734596"
spans list[struct] Character start/end boundaries and entity tags [{"start": 23, "end": 39, "label": "PERSON"}, {"start": 51, "end": 69, "label": "EMAIL"}, {"start": 72, "end": 80, "label": "PHONE"}, {"start": 87, "end": 97, "label": "SSN"}]
tokens list[string] Tokenized word representations ["DON'T", "SHACE", "buX", ...]
labels list[int64] BIO token classification tags [0, 0, 0, 1, 2, ...]
input_ids list[int32] Pre-encoded RoBERTa subword IDs [0, 15309, 6, ...]
attention_mask list[int8] Self-attention mask [1, 1, 1, ...]

Supported Entity Types

PERSON, EMAIL, PHONE, SSN, ADDRESS, CREDIT_CARD, PASSPORT_NUMBER, DATE_OF_BIRTH, IP_ADDRESS, USERNAME, PASSWORD.

4. UM-DLP Public Benchmarking Dataset

Schema Definition

Field Type Description Sample Value
ID int64 Test case identifier 1
Category string Data domain tested "Personal Identifiable Information (PII-Financial related)"
Type string Attack or payload category "Positive Direct", "Positive Obfuscated", "Negative Keyword"
Test data string Exact text prompt sent to DLP engine "Please help me verify if my credit card number 6011-2398-5674-9012 has been comp..."
Ground Truth (sensitive/non sensitive) string True binary sensitivity label "sensitive" or "non sensitive"
UM MAISON Detection string Reference system prediction "sensitive"
Your detection (sensitive/non sensitive) float/null Evaluation placeholder null

5. Contextual Sensitive Data

Schema Definition

Field Type Description Sample Value
column_name string Database or table attribute name "condition"
records string Sample values extracted from column "['0', 'active', 'active', '1', '1']"
instruction string System instruction prompt "You are a PII classification system. Given a column name and records, determine..."
input string Formatted prompt input "Column name: condition\nRecords: ['0', 'active', 'active', '1', '1']"
output string Ground-truth sensitivity classification & reasoning "Non-Sensitive (Categorical system state)"

6. Enron Corporate Email Corpus (AESLC)

Schema Definition

Field Type Description Sample Value
subject string Email subject header "Service Agreement"
body string Full corporate email body "Greg/Phillip, Attached is the Grande Communications Service Agreement. The business points can be f..."
label string Sensitivity classification "Internal"
source string Data provenance "Enron_Corporate"

7. STARGATE CIA Scanned PDF Archive

  • Local Path: ai-service/data/raw/stargate_pdfs/
  • Benchmark Target: ai-service/data/benchmarks/ocr_pipeline/pdf_test_corpus/
  • Modality & Format: Scanned PDF files (~300+ MB)
  • Volume: 7,394 scanned PDF documents
  • Origin / Provenance: GotThatData/STARGATE / CIA CREST declassified PDF archive (Grill Flame, Center Lane, Sun Streak, Star Gate).
  • Compliance & Threat Model: OCR text extraction, scanned image leak detection, noisy document binarization, declassified archive redaction verification.

Characteristics

  • Format: Scanned multipage PDFs (300 DPI, greyscale/bilevel, 1970s–1990s typewriter font).
  • Visual Artifacts: Redaction black-bars, official stamps (TOP SECRET, DECLASSIFIED), handwritten marginalia, degraded typewriter text.

🔬 Model Training & Split Summary

The primary 4-tier document classifier in ai-service/data/scripts/train_classifier.py was trained on the processed splits:

Class Partition Breakdown 

Total Records: 7,615
├── 70% Train:      5,330 rows (Confidential: 3725, Internal: 840, Public: 664, Highly Confidential: 101)
├── 15% Validation: 1,142 rows (Confidential: 798,  Internal: 180, Public: 142, Highly Confidential: 22)
└── 15% Test:       1,143 rows (Confidential: 799,  Internal: 180, Public: 142, Highly Confidential: 22)

Baseline Model Evaluation Metrics 

                     precision    recall  f1-score   support

       Confidential       0.99      0.91      0.95       799
Highly Confidential       0.47      0.82      0.60        22
           Internal       0.99      0.99      0.99       180
             Public       0.71      0.96      0.82       142

           accuracy                           0.93      1143
          macro avg       0.79      0.92      0.84      1143
       weighted avg       0.95      0.93      0.93      1143
Downloads last month
35
Total file size:
10.8 GB