Sigurdur/pii-masking-mim-gold

PII Masking Benchmark for Icelandic (MIM-GOLD-NER, augmented)

A training/benchmark set for Icelandic PII detection, built by taking all 3 splits of vesteinn/icelandic-ner-MIM-GOLD-NER (train, validation, test — same sentences, same split boundaries) and layering PII-specific annotations on top of its existing named-entity gold labels. It was created to benchmark privacy_filter, a rule-based Icelandic PII masking tool, but the annotations are generic and usable for any Icelandic PII-detection work.

Only test is used to benchmark privacy_filter (5,890 sentences; see Benchmark results below) — train/validation carry the identical annotation but aren't scored against, they're published so this dataset has the splits a NER-shaped dataset is expected to have, without implying privacy_filter was evaluated against more than it actually was.

split	sentences
train	46,454
validation	6,480
test	5,890 (the only split used for scoring)

Schema

Each row (across all 3 splits) is one sentence:

field	type	description
id	string	the sentence's id in its original MIM-GOLD-NER split
text	string	the sentence, detokenized from MIM-GOLD-NER's token list back into raw text (see Derivation below)
ner_entities	list of {start, end, label, text}	MIM-GOLD-NER's original, unmodified, human-annotated gold entities, all 8 label types (Person, Location, Organization, Miscellaneous, Date, Time, Money, Percent), with the same [start, end) character offsets as text
pii_entities	list of {start, end, label, text, source}	the PII-relevant view of the sentence — see below

pii_entities merges two annotation layers, distinguished by source because they have very different evidentiary status:

• source: "gold" — the subset of ner_entities whose label maps cleanly onto a PII category, relabeled into that vocabulary: Person→person_name, Organization→company_name, Location→street_address. Still fully human-annotated gold, just renamed. (Note: MIM-GOLD's Location is broader than "street address" — it also covers cities, countries and buildings — so treating every Location as a street_address positive will look like it "misses" a lot if you evaluate a street-address-specific detector against it; that's a property of the source annotation, not an error in this dataset.)
• source: "rule_augmented" — spans found by running privacy_filter's own regex/checksum rules (kennitala, phone, email, bank_account, iban, credit_card, license_plate, ip_address, username) over the unmodified sentence text. MIM-GOLD-NER's 8-category scheme has no concept of any of these, so there is no independent human-annotated gold for them — this is the best available approximation, not verified ground truth. Concretely: 1,049 such spans were found across all 58,824 sentences (train 639, validation 295, test 115 — mostly phone and email), and manual spot-checking on the test split during benchmarking found real false positives in this layer — e.g. a fish-market price table row (154 100 140 421 58) Luhn-validating as a credit_card, and manuscript/architecture codes (AM 242, IA-32) matching the license_plate heuristic. Use this layer as "what a rule-based detector fires on in ordinary running text" (useful for estimating false-positive rate on non-PII prose), not as labels to train a model to reproduce blindly.

Derivation

1. MIM-GOLD-NER's own dataset-card loading script points at https://vesteinn.is/{train,dev,test}.txt, which returns 404 (dead as of 2026-09). The data was instead pulled from the Hub's auto-generated parquet mirror (vesteinn/icelandic-ner-MIM-GOLD-NER, revision refs/convert/parquet), which is content-identical.
2. MIM-GOLD-NER ships pre-tokenized (punctuation split into its own tokens). Sentences were detokenized back into raw text with a small heuristic joiner (space by default, no space before closing punctuation . , ) ] ? ! : ; « ' or after opening punctuation ( » [) that tracks exact per-token character offsets, which the BIO tags were then converted into gold spans against. This is not a general-purpose Icelandic detokenizer — it is only guaranteed self-consistent with the spans in this dataset. One known artifact: a handful of original annotations had already split things like mbl.is) oddly (e.g. into mbl.i + s + )), which the detokenizer reproduces as mbl.i s) rather than silently "fixing" — this shows up as a handful of malformed username matches (@mbl.i) in the rule_augmented layer, not a privacy_filter bug.
3. pii_entities was built as described above: the Person/Organization/Location subset of the original gold, relabeled, union'd with whatever privacy_filter's rules found for the 9 PII categories MIM-GOLD-NER doesn't annotate at all. Applied identically to all 3 splits.

Full source: benchmark/ in the privacy_filter repo (mim_gold_ner.py for loading/detokenizing, run_benchmark.py for scoring privacy_filter against this data, build_augmented_dataset.py for building and publishing this dataset). Rebuild with:

uv run python -m benchmark.build_augmented_dataset

Benchmark results

Computed on test only (5,890 sentences) — train/validation are not scored against, per project convention (see the split table above). Scoring privacy_filter's own predictions against the source: "gold" layer (strict = exact span match, lenient = any overlap, same label):

label	gold	pred	strict P/R/F1	lenient P/R/F1
person_name	1329	924	0.92 / 0.64 / 0.76	0.94 / 0.65 / 0.77
company_name	1389	48	0.81 / 0.03 / 0.05	0.98 / 0.03 / 0.07
street_address	1129	83	0.65 / 0.05 / 0.09	0.87 / 0.06 / 0.12

company_name/street_address recall is low by construction, not by bug: most MIM-GOLD Organization entities are foreign companies or organizations without a legal-form suffix (hf./ehf./...), which privacy_filter's Icelandic-focused detectors don't target, and most Location entities are bare place names (cities, countries) rather than street+house-number addresses.

License

MIM-GOLD (and therefore MIM-GOLD-NER, and therefore this dataset's text and ner_entities) is sampled from the Icelandic Gigaword Corpus, Part 1 (IGC1), distributed by CLARIN-IS / the Árni Magnússon Institute for Icelandic Studies under the IGC1 license. That license is free of charge for research use, requires attribution ("IGC1 should be referenced as a source in publications of research where IGC1 has been utilized. The same applies to the release of any language technology tools that have used IGC1."), and restricts further redistribution of the raw corpus text beyond what Icelandic Copyright Act No. 73/1972 art. 14 allows. This dataset is published following the precedent of vesteinn/icelandic-ner-MIM-GOLD-NER itself being openly published on the Hub. All intellectual property rights in the underlying text remain with the original text providers and the Árni Magnússon Institute; this repo adds only the pii_entities annotation layer and the relabeling of ner_entities.

If you use MIM-GOLD-NER (the entities/text this dataset builds on), cite:

@misc{20.500.12537/42,
 title = {{MIM}-{GOLD}-{NER} – named entity recognition corpus},
 author = {Ing{\'o}lfsd{\'o}ttir, Svanhv{\'{\i}}t and Gu{\dh}j{\'o}nsson, {\'A}smundur Alma and Loftsson, Hrafn},
 url = {http://hdl.handle.net/20.500.12537/42},
 note = {{CLARIN}-{IS}},
 copyright = {Icelandic Gigaword Corpus Part1},
 year = {2020} }