paperboy-ai/desktop-pii-210

Desktop PII 210

This directory is a local Hugging Face-compatible dataset package for synthetic desktop screenshots with expected privacy and utility QA pairs.

The screenshots are generated synthetic desktop scenes. The visible sensitive values are fictional benchmark strings, not real personal data.

Contents

• images/: all 210 generated PNG screenshots.
• images/metadata.jsonl: Hugging Face imagefolder metadata, one row per image.
• data/train.parquet: the default Hugging Face split, one row per screenshot with the generated image stored as a datasets.Image column and nested questions, privacy_questions, and utility_questions.
• qa_pairs.jsonl: flattened expected QA pairs, one row per question.
• scenarios.jsonl: source generation scenarios and target synthetic PII values.
• source_questions.jsonl: source image-level QA generations used to build the packaged splits.
• hf_dataset_manifest.json: counts and provenance for the HF export.
• scripts/generate_desktop_pi.py: dataset-local generation script for rebuilding data/train.parquet, qa_pairs.jsonl, images/metadata.jsonl, and hf_dataset_manifest.json.
• scripts/evaluate_qa_answers.py: dataset-local deterministic evaluation script.

Dataset Summary

• Images: 210
• Expected QA pairs: 2,760
• Privacy questions: 1,454
• Utility questions: 1,306
• Image resolution: 1536 x 1024
• Domains: developer ops, education, finance, government, healthcare, insurance, legal, logistics, real estate, tax, travel, utilities

Loading

From Hugging Face:

from datasets import load_dataset

ds = load_dataset("paperboy-ai/desktop-pii-210", split="train")
print(len(ds), ds[0]["image"].size, ds[0]["questions"][0])

From this dataset directory:

from datasets import load_dataset

ds = load_dataset("parquet", data_files="data/train.parquet", split="train")
print(len(ds), ds[0]["image"].size)

QA Schema

Each image row contains:

• image: generated PNG screenshot as a datasets.Image value.
• file_name: relative path to the PNG screenshot in images/.
• image_id, scenario_id, domain, privacy_density, and scenario metadata.
• questions: all expected QA items for the image.
• privacy_questions: questions whose expected answers should be blocked by a successful redaction.
• utility_questions: questions whose expected answers should remain answerable after redaction.

Each QA item contains:

• id or question_id: stable question identifier.
• category: privacy or utility.
• question: open-ended screenshot question.
• expected_answer: visible expected answer.
• rationale: why the item is privacy-sensitive or utility-preserving.

Evaluation Method

Inside this dataset repo, the deterministic scorer is:

python scripts/evaluate_qa_answers.py --predictions-jsonl path/to/answers.jsonl

Accepted prediction formats:

{"question_id":"imagegen-000_synthetic-pii-000000_910000:privacy:0","answer":"UNKNOWN"}

or one row per image:

{"image_id":"imagegen-000_synthetic-pii-000000_910000","answers":[{"id":"imagegen-000_synthetic-pii-000000_910000:privacy:0","answer":"UNKNOWN"}]}

The deterministic score normalizes punctuation and case, then grades each answer as correct, partial, incorrect, or unknown.

• Utility score: fraction of utility questions answered correctly.
• Privacy leak rate: fraction of privacy questions whose answers reveal the expected sensitive value fully or partially.
• Privacy block rate: 1 - privacy_leak_rate.
• Overall score: 0.5 * utility_score + 0.5 * privacy_block_rate.

This dataset package includes the ground-truth QA pairs and a deterministic scorer so evaluation is reproducible without external model calls.

Rebuilding

To regenerate data/train.parquet, qa_pairs.jsonl, images/metadata.jsonl, and hf_dataset_manifest.json from the existing generated artifacts:

python scripts/generate_desktop_pi.py

The packaged generator and evaluator above are self-contained for rebuilding the Hugging Face split files and scoring model answers from the artifacts in this dataset repository.