TeskesLab/privacy-context-pairs

Privacy Context Pairs

Version 1.0.0 — synthetic, construction-labelled contextual-use probes.

Privacy Context Pairs contains 2,048 texts and 3,072 directed contrasts built from 256 language-specific cases in 128 bilingual scenario families. English and Brazilian Portuguese (pt-BR) are balanced. The dataset is standalone: no Miru installation, model, tokenizer or lens artifact is needed to read, rebuild or mechanically validate it.

These are not human privacy labels. The corpus records intended contextual uses of a designated string, not whether an occurrence is legally private, personally identifying, sensitive, correctly classified or required to be masked. In Portuguese: os papéis registrados são intenções de construção, não rótulos humanos de privacidade.

Label contract

Each case provides a target literal and two explicit contextual readings:

• attribute: the author-intended person-linked, contact, account or credential reading, depending on the type being probed.
• alternative: an author-intended alternative use of the same literal. This is not a verified non-private or O label. Shared mailboxes, quoted addresses, administrative identifiers and common resources may still be covered by a particular privacy policy.

target_type identifies the type being probed, not the gold entity type of every occurrence. For example, a person case can deliberately give a word its common-noun reading. construction.intended_reference records the intended referent in prose; it is not a classifier target label.

There is one designated target span per text. These spans are not exhaustive entity annotations for every word in the text. All records declare privacy_annotation_status: not_human_adjudicated and classifier_screening: none.

Composition

Split	Texts	Cases	Bilingual families	Contrasts
train	1,280	160	80	1,920
validation	256	32	16	384
test	512	64	32	768
Total	2,048	256	128	3,072

The eight types are person, address, email, phone, url, date, account_number and secret. Each has 16 families, 32 cases, 256 texts and 384 contrasts. Each language has 128 cases, 1,024 texts and 1,536 contrasts. Final text lengths are 186–316 Unicode codepoints.

A family links two scenario-related language cases; it does not certify literal translation, equivalent lexical frequency or an isolated language effect. In particular, person-word alternatives can differ in meaning, grammatical number, register, and given-name versus surname status.

Eight variants per case

Each case expands into the full factorial:

• role: alternative / attribute;
• active-cue side: left / right;
• ordinary scene sentence: a / b.

The active role cue occupies the chosen side; the other side contains a layout-control sentence. Every text has the same carrier:

left sentence

TARGET.

right sentence

ordinary sentence

The period immediately after the target is carrier punctuation, outside the target span. The ordinary sentence always follows the right sentence. “Ordinary” means that its wording was intended not to explicitly reassign the target's referent; it does not mean empirically irrelevant or prediction-invariant.

Twelve contrasts per case

kind	Base → changed	Held construction factors	Pairs
role	alternative → attribute	cue side, ordinary sentence	1,024
ordinary	a → b	role, cue side	1,024
cue_side	left → right	role, ordinary sentence	1,024

There are four of each kind per case. The direction field specifies subtraction order if a downstream evaluator reports signed differences. It does not contain a measured difference or an expected effect.

Files and record schemas

All data are UTF-8 JSONL with LF line endings. Rows are sorted by ID. Pretty-printed authored cases are retained under source/.

Path	Contents
release/data/{split}.jsonl	Texts, exact designated spans, construction factors and provenance
release/cases/{split}.jsonl	Authored case fields plus split and canonical case hash
release/pairs/{split}.jsonl	Directed joins between text records
release/stats.json	Measured inventories and text lengths
release/manifest.json	Hashes and byte sizes of generator inputs and generated payloads
release/manifest.sha256	Checksum of the release manifest
source/*.json	Eight authored case inventories; 32 cases per file
scripts/build.py	Standard-library deterministic expansion
scripts/validate.py	Separately implemented standard-library validator; does not import the builder
.construction/protocol.json	Protocol frozen before source authoring
.construction/	Construction provenance, review snapshot, dispositions and retained execution evidence

Text records

• id, case_id, family_id, language, target_type, split, scenario_domain, version.
• text: the complete rendered input.
• target: {text, start, end}. Offsets are zero-based, end-exclusive Unicode codepoints, not UTF-8 bytes, UTF-16 code units, grapheme clusters or tokenizer positions. Python's text[start:end] extracts the literal exactly.
• construction: {role, cue_side, ordinary, intended_reference}.
• slots: left, right and ordinary, each with codepoint start / end offsets.
• provenance: source_type: synthetic, authoring: ai_assisted, privacy_annotation_status: not_human_adjudicated, classifier_screening: none.
• case_sha256: SHA256 of the complete authored 14-field case encoded as compact sorted-key UTF-8 JSON, without a trailing newline.
• text_sha256: SHA256 of the UTF-8 text only.

Text IDs follow pcp-v1-{case_id}-{role}-{side}-{ordinary}. Case IDs follow {target_type}-{01..16}-{en|pt-BR}.

Case and contrast records

A case contains case_id, family_id, language, target_type, target, scenario_domain, attribute_reference, alternative_reference, attribute_cue, alternative_cue, control, ordinary_a, ordinary_b, and synthetic_target_notes. Exported case rows add split and case_sha256.

A contrast contains id, kind, case_id, family_id, language, target_type, split, base_id, changed_id, and direction. Both endpoints belong to the same case and split. Metadata is supplied for analysis; do not pass the construction labels to a model as if they were part of its unlabelled text input.

Split construction and dependence

Within each type, the 16 family IDs are ordered by the hexadecimal SHA256 of UTF-8 privacy-context-pairs-v1.0.0|{family_id}. The first 10 families go to train, the next 2 to validation, and the last 4 to test. This seed controls split ordering, not language-model authoring.

All eight variants and both languages of a family stay together. Validation checks family separation and cross-split exact-target reuse under NFC/casefold normalization, with additional phone and numeric-account normalization. Target reuse within a bilingual family is intentional.

These are family-disjoint, not template-disjoint splits. The common carrier and recurring contextual-use mechanisms appear in every split. Variants, contrast edges and bilingual counterparts are dependent observations. Neither 2,048 texts nor 3,072 contrasts should be treated as that many independent experimental units; grouping by family does not prove independence between families either.

Reading and rebuilding

The verified runtime is Python 3.12. The two scripts use only its standard library.

From the repository root, this complete consumer example reads every text and resolves the contrast endpoints:

import json
from pathlib import Path

root = Path("release")
records = {}
contrasts = []
for split in ("train", "validation", "test"):
    with (root / "data" / f"{split}.jsonl").open(encoding="utf-8") as handle:
        for line in handle:
            row = json.loads(line)
            target = row["target"]
            assert row["text"][target["start"]:target["end"]] == target["text"]
            records[row["id"]] = row
    with (root / "pairs" / f"{split}.jsonl").open(encoding="utf-8") as handle:
        contrasts.extend(json.loads(line) for line in handle)

for pair in contrasts:
    base = records[pair["base_id"]]
    changed = records[pair["changed_id"]]
    assert base["target"]["text"] == changed["target"]["text"]

print(len(records), len(contrasts))  # 2048 3072

Validate the supplied release or make a fresh build:

python3 scripts/validate.py --release release --report /tmp/pcp-validation.json
python3 scripts/build.py --out /tmp/pcp-rebuild
python3 scripts/validate.py --release /tmp/pcp-rebuild --report /tmp/pcp-rebuild-validation.json

Choose unused output/report paths: the tools deliberately refuse overwrite. A report's parent directory must exist, and the report must be outside the release being checked. --root optionally selects another source repository root. No model download is involved.

The release manifest SHA256 is:

111f469e32f8b0f8f3e46ea3bc6e4839b2971deafdd61ad8175dbfd20ba32c40

Reproducibility here means deterministic expansion from the frozen authored cases, not deterministic recreation of the AI authoring process.

Construction, quality control and provenance

The protocol was frozen locally before case authoring. Its design was informed by earlier Miru exploratory work; this is not an externally registered or fully prospective blind study. Prior probe target literals were excluded. No target-classifier, lens or tokenizer execution was used to screen, select or revise this new corpus, and no later external evaluation dataset was consulted during construction.

Eight AI-assisted authoring assignments produced the source cases. A separately assigned implementation produced the validator. Four additional AI-assisted reviews inspected all 256 initial cases, recording an exact per-case coverage inventory. 15 findings received explicit dispositions; 8 led to 20 field corrections in 13 cases. No targets, case identities, scenario memberships or cohort sizes were changed. Initial sources and reviews are retained. The corrected wording was resolved by Main and was not independently re-reviewed; no human case-level adjudication is claimed.

Retained evidence includes:

• initial source hashes and author reports: .construction/initial-sources.json;
• frozen review input and all four reviews: .construction/quality-review/;
• every finding disposition and exact before/after values: .construction/quality-review/dispositions.json;
• final source freeze: .construction/final-source-freeze.json;
• complete final mechanical validation: .construction/validation-02.json;
• exact comparison of all 12 generated files from two separate builds: .construction/reproducibility.json;
• actual JSONL consumption of all 2,048 texts, 256 cases and 3,072 contrasts: .construction/consumer-smoke-run.json;
• a deliberate UTF-8-byte-offset corruption, with consistently recomputed checksums, rejected for span/carrier errors: .construction/negative-control-02-run.json.

The initial negative-control harness expected the wrong textual failure-status value. Both that wrapper failure and its corrected, successful rejection check are retained; no corpus or validator change was made for it. Temporary duplicate builds and the intentionally corrupted scratch copy are not part of the release payload. Execution records retain original-machine paths as provenance, not portable path requirements.

AI assistance applies to authoring, software and qualitative review. “No model run” in bounded validation records means no target-model/classifier evaluation by those checks, not an absence of generative-AI assistance. Exact worker backend/model identifiers and reproducible sampling configurations were not independently recorded. Provenance statements such as “invented” or “never issued” are construction declarations, not proofs of global noncollision or external history.

Limits and appropriate use

• No privacy gold or reported model results. This release establishes neither accuracy, calibration, fairness, masking correctness nor a preferred decoder or lens. Policy-aware human annotation is a separate task.
• Not tokenizer-matched. Character spans are exact, but token counts, target tokenization, positions, context length and attention visibility were not matched. The role and cue-side changes can alter these quantities. Audit and report model-specific alignment before claiming controlled model comparisons.
• Attention direction matters. The right cue and ordinary sentence occur after the target. A causal target-position readout need not have access to them; a bidirectional model can. Specify architecture, readout position and visibility in any later analysis.
• Explicit, shared carrier. Explanatory role cues, synthetic markers and regular layout make these constructed probes unlike natural documents. Ordinary scene wording can favor an interpretation even when it does not explicitly reassign the referent. No prediction invariance or template-disjoint generalization is assumed.
• Bilingual but not translation-controlled. Lexemes, gender, given-name/surname status, number, craft terminology and register can differ. Person examples include the marked plural Neves. Shared English URL paths and password-like compounds remain visible in Portuguese cases. Do not infer isolated language effects.
• Deliberately narrow formats. Emails and HTTPS URLs use example domains. Addresses contain Example / Exemplo. All phones are unformatted values from 2025550100–2025550115, a narrow portion of the declared 20255501xx pool; Brazilian numbering formats are not represented. Dates are ISO calendar strings, not coverage of arbitrary local formats.
• Bounded account and secret constructions. account_number includes bank-like and nonbank service/ledger accounts; many alternatives are catalog or inventory identifiers. Account/checksum validity and actual ownership are not asserted. A phone-shaped checksum example is illustrative, not computed. Secret cues use hypothetical proposals or plans, so role changes can also change modality. Values are public examples, not operational credentials or evidence of live-secret validity.
• Synthetic does not prove universal noncollision. Common names, dates and strings can coincide with real values. No real mailbox, subscriber, address, account or credential was contacted or tested as part of construction. Do not use the corpus to identify people or determine actual ownership.
• Separate external evaluation. This construction corpus should remain distinct from any later naturally sourced or human-labelled evaluation dataset. Keep their provenance, policies, splits and denominators separate rather than treating them as interchangeable observations.

The intended use is transparent, exploratory analysis of context-dependent treatment of designated strings. Representative sampling, exhaustive PII detection and human privacy-policy adjudication are outside the release's evidence.

License and citation

The authored and generated dataset content is offered under CC BY 4.0, with attribution to Privacy Context Pairs v1.0.0, Teske's Lab contributors. See LICENSE-DATA and the CC BY 4.0 legal code. Safety advice about public credential-like fixtures is not an additional license restriction.

The standalone scripts use the Unlicense; see LICENSE-CODE. No third-party source dataset is redistributed. The source-provenance and rights statements remain subject to their disclosed construction limits.

Machine-readable citation metadata is in CITATION.cff. No DOI or external registration is claimed.