NagaYu/crowd-anonymity-sets

Crowd — anonymity-set sizes for attribute combinations

Read this first

This dataset describes nobody. Every row is generated from a probability model over attribute values built from published aggregate statistics. There is no person in it, no record to link, and no index that could be searched for an individual. The prose is synthetic.

The label is a head-count, not an identity. Each row's target is log10(number of people in the reference population matching this combination of attributes). The dataset exists to train a model that says how many people a description could be, never which person it is.

Counts are censored at a reporting floor of 5. Any combination matching fewer than five people is labelled "at most five" rather than with its exact value. A dataset that taught a model to distinguish "one person" from "two people" would be training a re-identification oracle.

It does not support a claim of anonymity. Information usable for re-identification is unbounded in principle. Models trained here estimate and prioritise; they do not certify.

What this is

Pairs of (natural-language description, exactly-computed anonymity-set size).

{
  "text": "都内で四十代前半の男性が子どもの心臓を診る医師として働いている。この道15年。",
  "evidence": {"region": "region:jp-13", "age": "age:40-44", "sex": "sex:male",
               "occupation": "occupation:physician_pediatric_cardiac_surgery",
               "experience": "experience:11-15"},
  "log10_count": 1.52,
  "log10_count_independent": -0.06,
  "independence_error_log10": 1.58,
  "log10_sigma": 0.46,
  "below_floor": false,
  "euphemism_fraction": 0.8,
  "profile": "journalism",
  "sampling": "rarity_tilted"
}

The label is not annotated, crowdsourced or estimated. It is computed by exact belief propagation over a tree-structured log-linear model fitted to published marginals — see How the label is computed below.

Why it exists

Text anonymisation today offers two options. Delete every named entity, which is cheap and destroys the document. Or ask a large language model to rewrite it, which works and does not fit on a phone. Both skip the question that actually determines risk: how many people does this description fit?

Answering that needs supervision that did not exist: descriptions paired with their true anonymity-set size. This dataset is that supervision.

Fields

field	type	meaning
text	string	the synthetic description
evidence	JSON string	normalised attribute values, family -> lattice node id
families	list[string]	which attribute families appear
n_attributes	int	how many
log10_count	float	the target. log10 head-count, censored at the floor
log10_count_uncensored	float	the raw value before censoring, for analysis only
count	float	10 ** log10_count
log10_count_independent	float	what the naive product-of-marginals assumption gives
independence_error_log10	float	log10_count − log10_count_independent
log10_sigma, log10_lo, log10_hi	float	reference-model uncertainty and 90% interval
below_floor	bool	the true count is at or below the reporting floor
mentions	JSON string	gold spans: character offsets, surface, attribute, register
task_spans	JSON string	spans of non-identifying task content
task_fields	JSON string	the fields a downstream reader needs
profile	string	clinical / legal / hr / journalism / generic
template_id	string	which surface realiser produced it
euphemism_fraction	float	fraction of mentions written obliquely
sampling	string	natural or rarity_tilted (see Sampling)
pair_id, pair_form	int, string	matched-pair id and register, in euphemism_probe
assumption_notes	list[string]	which modelling assumptions this row's count depends on

Splits

split	rows	purpose
train / validation / test	~216k / ~16k / ~16k	the regression task
euphemism_probe	~6k (3k pairs)	matched-pair robustness measurement

Splits are made on a hash of the attribute combination, not of the text. A random split would put "42歳の男性医師、東京" in train and "四十代前半の男性の医師、都内" in test, and the reported test error would be measuring paraphrase memorisation. Splitting on the combination means the test set contains crowds the model has never been told the size of.

euphemism_probe holds matched pairs: the same attribute combination rendered once plainly and once obliquely, sharing a pair_id. The difference in a model's prediction between the two halves is a direct measurement of register robustness with the confound held fixed.

How the label is computed

The reference population is a tree-structured log-linear model over ten attribute families:

p(x) ∝ Π_i φ_i(x_i) · Π_(i,j)∈E λ_ij(x_i, x_j),     E a forest

• φ_i are fitted by iterative proportional fitting (raking) so that every one-way marginal of the model reproduces its published counterpart to 1e-7. This is the same construction statistical agencies use for disclosure control.
• λ_ij encode published two-way dependencies (a 22-year-old is not a consultant surgeon; nurses are overwhelmingly female on the register; foreign residents concentrate in three prefectures).
• Because E is a forest, the partition function and every conditional count are computable exactly by one sum-product sweep. No sampling, no variational approximation. A count takes ~0.1 ms.

Ten families, with generalisation lattices: age (5-year bands → decades → coarse), sex, region (47 prefectures → 8 blocks → country), locality, occupation (6 levels deep, from 小児心臓外科医 up to 専門的・技術的職業), education, household, nationality, condition (including rare diseases), experience.

What the independence assumption costs

Every row carries independence_error_log10. Over the corpus the mean absolute error of the naive product-of-marginals assumption is 0.38 log10 (a factor of 2.4), and the worst case exceeds 13 orders of magnitude. Any tool that multiplies marginals is making an error of that size silently.

Sampling

Roughly 55% of rows are drawn from the natural population distribution and 45% are rarity-tilted: the long-tailed families (occupation, condition, region) are resampled uniformly over their leaves. Every row records which.

Why: sampling only from the population distribution produces a corpus whose median document already hides in a crowd of ~10⁶, where every masking decision is trivial. The tilted fraction fills the 10⁰–10⁴ range where the decisions actually happen. The label stays exact either way — the counting engine does not care how the combination was chosen — so this changes coverage, not correctness. Report natural-distribution and tail performance separately.

Resulting coverage (250k rows): every decade from 10⁰ to 10⁸ holds between 0.6k and 51k documents; 13.6% of rows sit at or below the reporting floor.

Registers

Every attribute is realised sometimes plainly and sometimes obliquely:

plain	oblique
42歳	四十代前半 / 四十二歳 / さん(42)
小児心臓外科医	子どもの心臓を診る医師 / 先天性心疾患の手術を担当する医師
東京都	都内 / 23区
心臓移植を受けた	新しい心臓をもらった / 移植で命をつないだ
単身世帯	ひとりで暮らしている

Predicate-form euphemisms (「持病がある」,「仕事をしていない」) appear as whole clauses with no noun for an entity tagger to latch onto — the hardest case, and the one a corpus of noun phrases would miss entirely.

Limitations — read before quoting a number

1. The reference bundle is a research fixture. Its figures are transcribed from published Japanese statistics and tagged published / approximate / derived / illustrative per entry. Rows tagged approximate are believed to match the published series but have not been re-verified against the primary table; interaction magnitudes tagged illustrative are qualitative modelling choices, not measurements. Do not cite the absolute numbers. Re-run scripts/fetch_reference.py against authoritative extracts (the e-Stat table ids are in the script) before using this for a real disclosure decision.
2. The text is synthetic and template-generated. Vocabulary and syntax are narrower than real clinical or legal prose. Use TAB or another real benchmark for span-level evaluation; this corpus is for the combination-risk task, which no span-annotated corpus supplies.
3. Closed-world extraction. The generator and the shipped extractor share the reference bundle's surface lists. Any round-trip accuracy measured on this corpus is a coverage measurement, not a robustness measurement. This is the main reason to train a model on the text rather than rely on the lexicon.
4. One value per family. A person may hold several diagnoses; the model carries the most disclosive one. Multi-morbid descriptions are therefore over-estimated — Crowd will think the crowd is larger than it is.
5. National, time-averaged marginals. No migration, no cohort effects, no sub-prefecture geography.
6. Japan only. The lattices, the surfaces and the statistics are Japanese. The bundle format is general; another country needs its own.
7. Sensitive attributes are present (condition, nationality) because they are genuine quasi-identifiers and omitting them would understate risk. They appear only as aggregate distributions.

Links

• Code, evaluation and figures: github.com/NagaYu/crowd
• Model trained on this dataset: NagaYu/crowd-anon-0.1b

Reproduction

python scripts/fetch_reference.py      # compile the reference bundle from published figures
python scripts/build_dataset.py --n 250000 --out data/build

Fully deterministic given --seed. No network access required.

Citation

@misc{crowd2026,
  title  = {Crowd: estimating anonymity-set size from attribute combinations},
  year   = {2026},
  note   = {Dataset and model. Reference statistics compiled from Japanese
            government publications under 政府標準利用規約(第2.0版).}
}

Source attribution for the reference statistics is required by their licence and is listed in the bundle's provenance block: 総務省統計局『令和2年国勢調査』・ 『労働力調査』, 厚生労働省『医師・歯科医師・薬剤師統計』・『衛生行政報告例』, 日本弁護士連合会ほか各士業団体, 日本臓器移植ネットワーク, 出入国在留管理庁 『在留外国人統計』.