inovruzova/polygraf-ner-stage2-expanded

Polygraf / NER Stage 2: Expanded Training Dataset

The Stage 2 output of the Polygraf Applied NLP / NER technical task. It combines the 95 corrected records from Stage 1 with 505 new records drawn from public English corpora and relabeled under the annotation policy reproduced in full below.

The Stage 1 dataset is not modified or overwritten. It is published separately at inovruzova/polygraf-ner-stage1-cleaned.

Summary

Item	Value
Records	600
From Stage 1	95
New in Stage 2	505
Train / test	481 / 119
Total labeled spans	3,237
Language	English
Labels	PERSON, ORGANIZATION, LOCATION, TIMEDATE, PRODUCT, WORKOFART, JOB, AMOUNT

Record schema

• unique_index (int): record identifier
• source_text (string): the sentence or passage
• privacy_mask (list): labeled entities, each with start, end, label, value
• source (string): which corpus the passage came from
• license (string): the licence governing that passage

Every record satisfies the following, checked by assertion at export:

• every label is one of the eight declared labels
• source_text[start:end] == value for every span
• no span carries leading or trailing whitespace
• no two spans within a record overlap
• every unique_index is unique

Data sources

This dataset is a collection. Each portion keeps the licence of the corpus it came from, and those licences are not identical. The license field on every record states which one applies.

Source	Licence	Register	Records
rjac/kaggle-entity-annotated-corpus-ner-dataset (GMB)	ODbL 1.0 / DbCL 1.0	news wire	225
DFKI-SLT/few-nerd, supervised	CC BY-SA 4.0	encyclopedic	280
polygraf-ai/applied-nlp-ner-candidate-starter-100 via Stage 1	evaluation-use-only	mixed	95

GMB provenance: Groningen Meaning Bank, then the Kaggle extract, then this mirror. ODbL and CC BY-SA are both share-alike, and those obligations carry forward to the portions derived from them.

An occupational title list, gpriday/job-titles, was used to flag JOB spans for review. It is 65,248 titles merged from ESCO (EU), O*NET v29.3 (US Department of Labor) and OSCA (Australia). No text from it appears in this dataset.

ai4privacy/pii-masking-openpii-1m (CC BY-4.0) was cleared on licensing but excluded at ingest for a practical reason: the corpus is a single 4.61 GB JSONL file ordered by language, and sequential streaming did not reach the English rows in the time available.

Labels the sources do not annotate

Neither source annotates JOB or AMOUNT. GMB tags no quantities, and Few-NERD deliberately omits numerical, value and date types. So six of the eight labels here are corrections of existing source annotations, and two are additions: all 263 JOB and 244 AMOUNT spans in the new records were introduced by a model, JOB by GLiNER and AMOUNT by spaCy en_core_web_lg, then corrected under the policy.

For JOB this is a constraint of the field rather than a shortcut. No public English NER corpus provides occupational title annotation; a survey of the juand-r/entity-recognition-datasets catalogue found none, and the ai4privacy releases use TITLE for honorifics rather than professional roles.

OntoNotes was considered as an AMOUNT source, since its label set covers seven of the eight labels here, and rejected on licensing. The LDC user agreement grants no right to copy, redistribute or publish the data outside the signatory's own research group, which a public repository containing the source text would violate, and the Hugging Face mirrors declare their licence as "other". spaCy's weights are MIT-licensed and contain no OntoNotes text.

Label distribution

Label	Stage 1 (95)	Stage 2 new (505)	Combined (600)
PERSON	106	415	521
TIMEDATE	109	515	624
LOCATION	79	403	482
ORGANIZATION	96	306	402
JOB	65	263	328
AMOUNT	80	244	324
PRODUCT	84	216	300
WORKOFART	54	202	256
Total	673	2,564	3,237

How the labels were produced

1. Source annotations mapped onto the eight-label schema. Tags with no target label are dropped and counted: 2,230 spans across 20 tag types, almost all nationality adjectives, named events and Few-NERD other subtypes, which the policy places outside the schema.
2. spaCy adds candidate AMOUNT spans and GLiNER adds candidate JOB spans. 315 AMOUNT and 207 JOB candidates were added.
3. A policy linter applies every mechanically decidable rule. 138 changes.
4. An LLM (gemini-3.6-flash, temperature 0) applies the full policy to each passage, receiving the current annotation as inline XML tags and returning corrected tags. Offsets are recovered by removing the tags and verifying that the remaining text is byte for byte identical to the input.
5. The linter runs again. 11 changes, all function word edges.
6. Remaining suspicions are ranked into a review queue.

The LLM pass covers 500 of the 505 new records. Four remain from an earlier partial run and one was rejected for altered text. Those five keep their pre-LLM annotation, which is the designed failure mode rather than a crash.

How the labels were validated

Assertions at export, listed under Record schema. These catch structural corruption, not annotation error.

An audit of 40 records drawn at random from the 505 new records and marked by hand. Stage 1 records were excluded because they were reviewed during Stage 1. 172 labels were checked: 169 correct, 3 wrong, 2 entities missed entirely. Span precision is 98.3 percent. Recall is approximately 98.8 percent, but one record's missed count was left blank, so that figure is indicative.

All findings were three problems: a nationality adjective left inside a JOB span against R28, twice; a currency symbol included without a preceding approximator against R4, once; and two misses, an unlabeled JOB (founder in "Microsoft founder Bill Gates") and an unlabeled ordinal.

A review queue of 117 linter flags, sampled rather than worked. Five flags of each kind were reviewed by hand. The JOB-list check was right 1 time in 5, and found that one error for the wrong reason. The numeral check was right 0 times in 5, but its persistent firing on ordinals is what exposed the policy gap that R33 closes, so it is retained unchanged. Neither check should be described as validation; JOB in particular has no independent verification anywhere in the pipeline.

Train and test split

Records are grouped by source, by the rarest label they contain, and by length band, then each group is split 80/20. Seed 202608. No validation split, following the task instruction.

Split	Records	Spans	Spans per record	Sources
All	600	3,237	5.39	stage1 95, gmb 225, few-nerd 280
Train	481	2,637	5.48	stage1 77, gmb 181, few-nerd 223
Test	119	600	5.04	stage1 18, gmb 44, few-nerd 57

Every label appears in both halves. Of the 33 records longer than 500 characters, 6 fall in test, 18 percent against a target of 20.

An earlier version drew the test split only from human-verified records, reasoning that a test split containing auto-labeled records measures agreement with the pipeline rather than correctness. That version was rejected: the verified pool is almost entirely Stage 1 records, which carry a different register mixture and much higher entity density, so the result was a test split with no register overlap with training at all, which would have depressed downstream metrics for reasons unrelated to model quality.

The trade-off is stated rather than resolved. The test split can be clean or representative but not both, and representativeness was chosen. Test metrics should be read as agreement with a reference measured at 98.3 percent precision rather than as ground truth. The source field lets anyone re-derive a verified-only split.

Known issues

• JOB and AMOUNT have no independent human-annotated source anywhere in the pipeline.
• The audit covers 40 of 505 new records. The rest are checked only by the linter and the export assertions.
• The review queue was sampled to measure the checks, not worked.
• Label and register are correlated. LOCATION and PERSON come disproportionately from the news wire corpus, PRODUCT and WORKOFART from the encyclopedic one, so a model may learn register cues rather than entity cues for those labels.
• Topic is not measured or balanced. The news portion skews heavily towards international politics.
• Only two registers are present besides the mixed Stage 1 material.
• Few-NERD's other-god subtype is left unlabeled, but no rule settles whether a named deity falls under R27. No record in this selection exercises it.

---

Labels

• PERSON: a named person, including given names and surnames.
  • Example: <PERSON>Barack Obama</PERSON> / <PERSON>Sarah Chen</PERSON>.
• ORGANIZATION: a named company or institution.
  • Example: <ORGANIZATION>Google</ORGANIZATION> / <ORGANIZATION>Mayo Clinic</ORGANIZATION>.
• LOCATION: a named place such as a city, country, region, street, landmark, or geographic area.
  • Example: <LOCATION>Berlin</LOCATION> / <LOCATION>New York City</LOCATION>.
• TIMEDATE: an expression that places something on a timeline, including dates, clock times, and durations used as time.
  • Example: <TIMEDATE>March 15, 2024</TIMEDATE> / <TIMEDATE>50 minutes</TIMEDATE>.
• PRODUCT: a named commercial product, device, or branded good.
  • Example: <PRODUCT>iPhone 15</PRODUCT> / <PRODUCT>MacBook Pro</PRODUCT>.
• WORKOFART: a named creative or published work such as a book, film, song, or titled publication.
  • Example: <WORKOFART>Oppenheimer</WORKOFART> / <WORKOFART>Spider-Man: Brand New Day</WORKOFART>.
• JOB: an occupational title or formal work role when it functions as such in the sentence.
  • Example: <JOB>software engineer</JOB> / <JOB>CEO</JOB>.
• AMOUNT: a measurable or countable quantity, not a time or date expression.
  • Example: <AMOUNT>50</AMOUNT> tickets / <AMOUNT>120,000</AMOUNT>.

Baseline labeling rules

These are the main rules. They are fixed and were not edited. Corrections and labeling decisions follow these main rules plus the added policy rules below.

• A labeled mention should cover the entity itself, not the surrounding grammar.
  • Example: Correct: She visited <LOCATION>Paris</LOCATION> yesterday. Incorrect: labeling visited Paris yesterday as one span.
• Articles, prepositions, conjunctions, and other function words should stay outside the span unless they are truly part of the proper name.
  • Example: Correct: He works at <ORGANIZATION>Google</ORGANIZATION>. Incorrect: <ORGANIZATION>at Google</ORGANIZATION>.
  • Example: <ORGANIZATION>The New York Times</ORGANIZATION> should keep The when it belongs to the established name.
• Bare type-words and category descriptors, meaning words that name a category rather than a specific entity, should not be labeled on their own. Examples include "person", "company", "organization", "team", "hospital", "city", "product", "book", and "quantity".
  • Example: The company hired 200 people. should leave company unlabeled.
  • Example: She joined <ORGANIZATION>Acme Corp</ORGANIZATION>. should still label the named organization.
  • Example: The person who called did not leave a name. should leave person unlabeled.
• A multi-word name should be one span when it forms one real named entity; separate entities should be labeled separately.
  • Example: <PERSON>Barack Obama</PERSON> should be one span.
  • Example: <PERSON>Barack</PERSON> and <PERSON>Michelle</PERSON> should be two spans.
• Coordinated names should be separate spans, unless the conjunction is part of one established name.
  • Example: <ORGANIZATION>Google</ORGANIZATION> and <ORGANIZATION>Microsoft</ORGANIZATION> should be two spans.
  • Example: <ORGANIZATION>Johnson & Johnson</ORGANIZATION> should stay one span because & belongs to the name.
• Ordinary punctuation and stray whitespace should stay outside spans, unless the punctuation belongs to the name or abbreviation.
  • Example: He moved to <LOCATION>Berlin</LOCATION>. should leave the final period outside the span.
  • Example: <ORGANIZATION>AT&T</ORGANIZATION> should keep & inside the span.
• Possessive markers should stay outside the span unless the full possessive form is the name.
  • Example: <PERSON>Maria</PERSON>'s laptop should leave 's outside the span.
• Labels should follow context, not surface form alone. The same words may be an organization in one sentence and a place in another.
  • Example: She works at <ORGANIZATION>Cambridge University</ORGANIZATION>.
  • Example: The conference was held in <LOCATION>Cambridge</LOCATION>.
• Quantities used as time should be labeled as TIMEDATE, not AMOUNT.
  • Example: She bought <AMOUNT>50</AMOUNT> tickets.
  • Example: The train arrives in <TIMEDATE>50 minutes</TIMEDATE>.

Added policy rules

These rules close gaps the baseline rules do not settle. They do not modify the baseline. Each is checked against the baseline and against the other added rules.

R1 to R32 were added in Stage 1. R33 was added in Stage 2 and is grouped by topic rather than by number, so the numbering in section A is not sequential.

A. Quantity and time boundaries

R1. Only value-changing modifiers go inside the span. Modifiers that change the value denoted are included. Modifiers expressing the writer's stance, or positioning the value relative to a reference point, stay outside. <AMOUNT>about 50</AMOUNT> and <TIMEDATE>at least two years</TIMEDATE> include the modifier; only <TIMEDATE>two years</TIMEDATE> and <TIMEDATE>two years</TIMEDATE> away do not, because "only two years" still denotes exactly two years.

R2. Approximators and comparators are inside the span, whether they precede or follow the value: about, approximately, around, roughly, nearly, almost, at least, at most, more than, less than, over, under, up to, ~. <AMOUNT>about $784,700.78</AMOUNT>, <TIMEDATE>less than six hours a night</TIMEDATE>.

R3. Units stay outside; magnitude multipliers stay inside. Units of measurement and currency codes stay outside because they name what is measured. Multipliers (k, m, bn, million, billion) and percent stay inside because they change the value. <AMOUNT>25000</AMOUNT> square feet, <AMOUNT>884k</AMOUNT> Tunisian Dinar, <AMOUNT>3.5%</AMOUNT>.

R4. Currency symbol contiguity exception. Where an approximator precedes a number carrying a prefixed currency symbol, the span extends across the symbol to remain contiguous: <AMOUNT>about $150</AMOUNT>. The symbol is included only when it lies between two parts of the span, never on its own.

R5. Numeral-derived quantity words are AMOUNT even without an exact value: <AMOUNT>thousands</AMOUNT>, <AMOUNT>several hundred</AMOUNT>. A bare quantifier with no numeral to anchor it stays unlabeled: a few, several, many.

R6. Date ranges. Ranges with a connective (from X to Y, between X and Y) are two TIMEDATE spans with the connective outside. Ranges written with a dash and no connective are a single span: <TIMEDATE>2019-2021</TIMEDATE>.

R33. Ordinals are AMOUNT. An ordinal names a position in a counted sequence, and a position is countable: the <AMOUNT>11th</AMOUNT> named storm, the <AMOUNT>third</AMOUNT> soldier, <AMOUNT>4th</AMOUNT>. An ordinal inside a date expression is TIMEDATE, under the baseline rule that quantities used as time are TIMEDATE: <TIMEDATE>the 4th of July</TIMEDATE>. An ordinal that orders events rather than counting things, or that is frozen in a set phrase, is not labeled under R29, because neither denotes a quantity: replaced its rifles first with British Lee-Enfields, first lady, first aid.

B. What counts as TIMEDATE

R7. TIMEDATE requires a determinate point, date, or duration. Unbounded expressions are not labeled, however phrased: always, never, ever, now, yet, already, initially, at any hour of the day, round the clock, in the course of. The test is whether the expression restricts the timeline.

R8. Frequency expressions require a determinate interval. <TIMEDATE>every Saturday</TIMEDATE>, <TIMEDATE>twice a year</TIMEDATE> and <TIMEDATE>24-hour</TIMEDATE> qualify. Every time, whenever, every now and then do not.

R9. Deictic day-words (today, tonight, yesterday, tomorrow) are TIMEDATE when they locate a specific event relative to the text's moment, and unlabeled when they mean the present era generally. Test: can the word be swapped for a calendar date without changing the meaning?

R10. Named calendar observances, holidays and seasons are TIMEDATE. <TIMEDATE>Remembrance Day</TIMEDATE>, <TIMEDATE>Easter</TIMEDATE>, <TIMEDATE>Ramadan</TIMEDATE>. An observance recurs at a fixed calendar position, which is what the TIMEDATE definition asks for.

R11. Bare numbers are TIMEDATE only when context marks them as a date or time: a preposition of time, a surrounding date expression, or an explicit temporal frame. Otherwise AMOUNT. Sharing <AMOUNT>3339</AMOUNT>, but in <TIMEDATE>2021</TIMEDATE>.

C. PRODUCT boundaries

R12. Vintage, model and edition years preceding a product name are a separate TIMEDATE span, not part of the PRODUCT span, unless the year is part of the registered name. <TIMEDATE>2012</TIMEDATE> <PRODUCT>Dom Perignon Brut</PRODUCT>, but <PRODUCT>Windows 95</PRODUCT>.

R13. Spec-like tokens (grades, sizes, counts, strengths) stay inside the PRODUCT span when part of how the product is named and sold, and are AMOUNT only when they modify from outside. Test: does the product exist under that name without the token? <PRODUCT>5 Puttonyos Royal Tokaji</PRODUCT>, but <AMOUNT>two</AMOUNT> bottles of <PRODUCT>Dom Perignon</PRODUCT>.

R14. Brand names stay inside the PRODUCT span when they form the leading component of a commercial name. The same name is ORGANIZATION only when it refers to the company as an entity in its own right: acting, employing, being located, being acquired.

R15. Trailing category descriptors stay outside the PRODUCT span, with or without an intervening comma. <PRODUCT>Dom Perignon Brut</PRODUCT>, Champagne, <PRODUCT>M282</PRODUCT> warhead.

R16. Software, programming languages, frameworks and query languages are PRODUCT. <PRODUCT>Apache Hadoop</PRODUCT>, <PRODUCT>SQL</PRODUCT>, <PRODUCT>Django</PRODUCT>.

R17. Video games are PRODUCT; films, books, songs and albums are WORKOFART. A title in both forms takes the label matching the referent in context: <PRODUCT>Final Fantasy VII</PRODUCT> (game) versus <WORKOFART>Advent Children</WORKOFART> (film).

D. ORGANIZATION and LOCATION

R18. Named commercial establishments are ORGANIZATION. Restaurants, hotels, casinos, shops. LOCATION is reserved for geographic and administrative places. A locative phrase does not convert an establishment to LOCATION.

R19. Possessive prefixes on dish preparations. Label the establishment and leave the preparation unlabeled: <ORGANIZATION>Andre's</ORGANIZATION> Classic Lobster Thermidor. A possessive prefix attributes a generic preparation; it does not convert it into a PRODUCT.

R20. Metonymy. Place names used for the institution they house remain LOCATION when the name is primarily a place; names that are primarily institutional remain ORGANIZATION even when used locatively. Momentum on <LOCATION>Capitol Hill</LOCATION>, but <ORGANIZATION>The Pentagon</ORGANIZATION> lifted a ban.

R21. Sports teams, clubs and leagues are ORGANIZATION, including when the name begins with a city or region. The geographic component stays inside the span and is not separately labeled: <ORGANIZATION>Buffalo Sabres</ORGANIZATION>.

E. PERSON and JOB

R22. Honorifics stay outside the PERSON span: Dr., Mr., Mrs., Ms., Prof., Sir, Rev. Dr. <PERSON>Amanda Martinez</PERSON>.

R23. Occupational words used as stand-in names are PERSON, not JOB. <PERSON>Guitarist</PERSON>: great player (a nickname for one performer) versus the <JOB>guitarist</JOB> was late (the role).

R24. Generic collective nouns for people are not JOB: officials, authorities, staff, employees, management, workers, residents, sources. A modifier can make one specific enough to qualify: <JOB>health officials</JOB>.

R25. Names of academic or professional fields are not JOB (psychology, finance, engineering). JOB requires a word denoting a person's role.

R26. Usernames and handles that denote a specific individual are PERSON: <PERSON>otavio.telin</PERSON>, <PERSON>@xbllygbsn</PERSON>. Opaque identifiers with no name component are unlabeled.

R27. Fictional characters and mascots are PERSON. <PERSON>Ronald McDonald</PERSON>, <PERSON>Cait Sith</PERSON>.

F. Cross-cutting

R28. Entities with no home label stay unlabeled: nationalities and ethnic groups, languages, and named events. Note the deliberate contrast with R10: a calendar observance marks a position on a timeline and fits TIMEDATE; a named event does not.

R29. Idiomatic and figurative uses are not labeled. A span is labeled only when it refers to an actual instance of the entity type. such May-December couples is unlabeled, since no month is referred to.

R30. Nested entities take the outer span only. The schema is flat and the training format admits no overlaps.

R31. Citation and document apparatus is not labeled: footnote markers, page and docket references, section symbols, internal cross-references.

R32. Named awards and prizes are WORKOFART.

---

Intended use and licensing

This dataset is a collection of passages from several corpora and does not carry a single licence. Each record's license field states the terms governing that record.

Attribution:

• GMB portion: Groningen Meaning Bank, via the Kaggle entity annotated corpus and the rjac Hugging Face mirror. Open Database License, Database Contents License 1.0.
• Few-NERD portion: Ding et al., Few-NERD. CC BY-SA 4.0.
• Stage 1 portion: Polygraf, applied-nlp-ner-candidate-starter-100, evaluation-use-only.

The share-alike obligations of ODbL and CC BY-SA apply to the portions derived from those sources. Anyone reusing this dataset should check the terms of the original sources for their specific use.

The annotations were produced during the Polygraf Applied NLP / NER technical task.

Reference

Klie, J.-C., Webber, B., and Gurevych, I. (2023). Annotation Error Detection: Analyzing the Past and Present for a More Coherent Future. The relabeling design in this stage reuses the Retag method described there, applied to found corpora rather than to the starter dataset.