blackpen-sk/parr-dataset

LOUPE — initial dataset

The init split contains all 300 original SynthPAI seed personas and all 7,823 original comments, grouped into one row per author. This is the initial source dataset: no anonymization, model inference, or predictions were performed.

Source and attribution

Adapted from SynthPAI by Yukhymenko et al., A Synthetic Dataset for Personal Attribute Inference. Source repository: eth-sri/SynthPAI, revision ba9edc9c8aa1d398bc43bb619a8d859351df9d9c. Dataset license: CC BY-NC-SA 4.0. Source paths, verified SHA-256 hashes, and author assignments are in manifest.json. The filtered 294-profile evaluation artifact is not used.

Changes: join comments to seed authors, convert the eight reference attributes to strings, render a deterministic eight-line profile, and add split/provenance metadata. Comment text and raw-file order are preserved exactly. Raw source files are cached unchanged by the builder; extra seed fields, usernames, generation instructions, reviewer labels, and historical model outputs are excluded from the normalized table.

Columns

• profile_id: original source author ID; rows follow seed-file order.
• profile: eight original reference values as strings (age, sex, residence, birthplace, education, occupation, relationship status, and income level).
• profile_text: deterministic labeled rendering of those eight values.
• comments: every original comment by this author in source order.
• usage_split: test for all 300 authors in every split; no development subset is reserved. This column is separate from the Hugging Face split names.
• provenance: source revision, profile value character intervals, and aligned comment IDs, one-based source line numbers, thread IDs, and parent IDs.

The structured profile and provenance are reference/evaluator data. Authors share threads, so author assignment does not eliminate thread dependence. This public synthetic corpus does not establish real-person anonymity or absence of model training contamination. Comments can disclose profile values directly.

from datasets import load_dataset

dataset = load_dataset("blackpen-sk/loupe-dataset", split="init")

Profiles and independently redacted comments

All nine splits preserve the original 300 authors, 7,823 comments, profile references, ordering and original profile outputs. comments_anonymized is a list aligned with comments and provenance.comment_ids. The original columns remain reference data and contain personal attributes.

Seven detectors process each comment independently using their pinned native labels, thresholds and decoding: Pencil, OpenMed Nemotron v2, OpenAI Privacy Filter, Perplexity PII, both GLiNER2 variants, and Presidio with spaCy. They receive only that comment, with no profile or ground truth. Detected source spans become [REDACTED]; overlaps are unioned and adjacent spans remain separate. Other text is preserved. Zero detections are successful unchanged outputs, not proof that no personal information remains. Token limits are checked without truncation; native long-text paths or overlapping windows preserve complete source coverage.

In init, the new column copies the original comments. In full_redaction, each entire comment becomes one [REDACTED] marker. This suppresses comment content while preserving comment count. Selecting the original comments column retains the historical full-profile-redaction condition. No inference or anonymity claim is implied by either reference. The union reader schema supplies null profile_anonymized values for init as before.

manifests/<split>.json retains profile lineage and points to comment manifests and provenance under comments/. Those files record pinned model/runtime identity, settings, raw detections, comment IDs, source hashes, spans, status, and output hashes. Original profile provenance remains under provenance/. Presidio uses the pinned standard English recognizers and en_core_web_lg on CPU, with no custom attribute rules. Its exact package and public-suffix-list identity is recorded instead of a Hugging Face model revision.

comment_release.json records source and release hashes. Earlier dataset revisions and experiments remain unchanged. No attacker or judge results are included. Detection coverage varies; redaction does not guarantee anonymity.

Both GLiNER runs also passed strict preprocessing replay on all 7,823 comments. The audit found no preprocessing failures, word truncation, or differences in encoder inputs between default and strict error handling. The audit is recorded in comments/validation/gliner_preprocessing.json.