Hasanblanka/polygraf-ner-stage4-final-dataset

Polygraf NER Stage 4 Final Dataset

This repository contains the final improved English NER dataset produced in Stage 4 of the Polygraf Applied NLP / NER technical task.

Dataset files

• data/train.jsonl: 1,954 training records.
• data/test.jsonl: 472 held-out test records.
• data/stage4_full.jsonl: all 2,426 records in one convenience file.
• split_provenance.jsonl: split and source-family provenance for every row.
• stage2_source_attribution.jsonl: original attribution for retained Stage 2 source rows.
• globalvoices_attribution.jsonl: article, author, license, source hash, and split metadata for all 1,549 Global Voices passages.
• stage4_synthetic_metadata.jsonl: generation groups and metadata for the 250 reviewed synthetic records.

The default Hugging Face configuration loads only train.jsonl and test.jsonl. The full file must not be loaded together with these splits for training because it contains both of them.

Each row follows the original character-span schema:

{
  "unique_index": 20250,
  "source_text": "Example English text.",
  "privacy_mask": [
    {"start": 0, "end": 7, "label": "PERSON", "value": "Example"}
  ]
}

All spans use exclusive end offsets. They are flat, sorted, non-overlapping, and satisfy source_text[start:end] == value.

Construction

Stage 4 Part 1 reviewed 506 existing training candidates, removed six body-less or objectively low-quality records, retained 500 natural records, and added 250 reviewed synthetic records targeting the error patterns found in Stage 3.

Stage 4 Part 2 collected 300 Global Voices articles, excluded 61 unsuitable or third-party republications, produced 1,642 candidate passages, and retained 1,549 passages after quality, duplication, prior-data overlap, and annotation QA. Article groups were kept intact during splitting.

Final split composition:

Component	Train	Test
Stage 4 Part 1 train	750	0
Frozen earlier test	0	127
Global Voices passages	1,204	345
Total	1,954	472

No earlier training record was moved into the test split. Exact train/test text leakage and Global Voices article-group leakage are both zero.

Label counts

Label	Train	Test
PERSON	1,743	415
ORGANIZATION	2,146	536
LOCATION	3,474	883
TIMEDATE	2,908	785
PRODUCT	494	121
WORKOFART	583	239
JOB	1,810	447
AMOUNT	2,551	711

Loading

from datasets import load_dataset

dataset = load_dataset(
    "Hasanblanka/polygraf-ner-stage4-final-dataset",
)
print(dataset)

Sources and licensing

This is a mixed-source dataset and must not be treated as having one permissive license for all rows.

• Starter records retain the evaluation-use-only terms in LICENSE.
• Retained Open Newswire records use their row-level CC BY or CC BY-SA terms; original publisher and URL metadata are stored in stage2_source_attribution.jsonl.
• Global Voices source passages are CC BY 3.0. Their article URLs, authors, license identifiers, and hashes are stored in globalvoices_attribution.jsonl.
• The 250 synthetic records were created for this task and their construction groups are documented in stage4_synthetic_metadata.jsonl.

Open Newswire is a discovery feed; rights remain with the listed original publishers. Global Voices attribution must be retained. See DATASET_LICENSES.md for the detailed source-family notes.

Limitations

• The data is English and news-heavy.
• Source and label distributions are not uniform.
• The annotation format is flat and does not support nested entities.
• WORKOFART, PERSON, and ORGANIZATION remain the most difficult labels for the final model.
• Synthetic records improve targeted coverage but do not represent naturally occurring frequency.
• This dataset is intended for evaluation and research in the technical task, not production deployment.

Full annotation policy

The complete policy in force for this Stage 4 dataset follows.

Annotation Policy

Status: final policy in force for the Stage 4 dataset. The baseline rules are unchanged; the additional rules were introduced during Stage 1 QA and retained through Stages 2–4.

Labels

• PERSON: a named person, including given names and surnames.
• ORGANIZATION: a named company or institution.
• LOCATION: a named place such as a city, country, region, street, landmark, building, or geographic area.
• TIMEDATE: an expression that places something on a timeline, including dates, clock times, ages, frequencies, and durations used as time.
• PRODUCT: a named commercial product, device, software item, or branded good.
• WORKOFART: a named creative or published work such as a book, film, song, article, report, or titled publication.
• JOB: an occupational title or formal work role when it functions as such in the sentence.
• AMOUNT: a measurable or countable quantity that is not a time or date expression.

Fixed Baseline Labeling Rules

These rules are fixed. Corrections follow these rules together with the extra policy rules below.

1. A labeled mention should cover the entity itself, not the surrounding grammar.
   • Correct: She visited <LOCATION>Paris</LOCATION> yesterday.
   • Incorrect: labeling visited Paris yesterday as one span.
2. Articles, prepositions, conjunctions, and other function words should stay outside the span unless they are truly part of the proper name.
   • Correct: He works at <ORGANIZATION>Google</ORGANIZATION>.
   • <ORGANIZATION>The New York Times</ORGANIZATION> keeps The because it belongs to the established name.
3. Bare type words and category descriptors should not be labeled on their own. Examples include person, company, organization, team, hospital, city, product, book, and quantity.
   • The company hired 200 people. leaves company unlabeled.
   • She joined <ORGANIZATION>Acme Corp</ORGANIZATION>. labels the named organization.
4. A multi-word name should be one span when it forms one real named entity; separate entities should be labeled separately.
   • <PERSON>Barack Obama</PERSON> is one span.
   • <PERSON>Barack</PERSON> and <PERSON>Michelle</PERSON> are two spans.
5. Coordinated names should be separate spans unless the conjunction is part of one established name.
   • <ORGANIZATION>Google</ORGANIZATION> and <ORGANIZATION>Microsoft</ORGANIZATION> are separate.
   • <ORGANIZATION>Johnson & Johnson</ORGANIZATION> stays together.
6. Ordinary punctuation and stray whitespace should stay outside spans unless the punctuation belongs to the name or abbreviation.
   • He moved to <LOCATION>Berlin</LOCATION>.
   • <ORGANIZATION>AT&T</ORGANIZATION> keeps &.
7. Possessive markers should stay outside the span unless the full possessive form is the name.
   • <PERSON>Maria</PERSON>'s laptop.
8. Labels should follow context, not surface form alone.
   • She works at <ORGANIZATION>Cambridge University</ORGANIZATION>.
   • The conference was held in <LOCATION>Cambridge</LOCATION>.
9. Quantities used as time should be labeled TIMEDATE, not AMOUNT.
   • She bought <AMOUNT>50</AMOUNT> tickets.
   • The train arrives in <TIMEDATE>50 minutes</TIMEDATE>.

Additional Policy Rules

A1. Temporal specificity, ages, and vague temporal adverbs

Calendar dates, clock times, seasons, durations, frequencies, ages, and relative/deictic expressions with a concrete temporal reference are labeled TIMEDATE. General adverbs that do not identify a time point or interval, such as ever, never, always, again, yet, just, initially, and already, are not labeled. Interrogative placeholders such as when and how long ask for a temporal value but do not provide one, so they are not labeled.

• Record 45: 19-year-old is an age and is labeled TIMEDATE.
• Record 72: Now and three years later place events on the timeline.
• Record 10: standalone ever, never, and before are not labeled.
• Record 58: How long asks for a duration but does not express one.

Reason: this prevents general discourse adverbs and question placeholders from inflating the time class while retaining real temporal references.

A2. Core quantity spans and unit boundaries

An AMOUNT span includes the numeric/count component, ranges, and scale words: 3.5, 3-4, 1.2 million, half, both, and thousands. Currency symbols and codes, measurement units, counted nouns, and approximation modifiers such as about, over, and at least stay outside.

Digits alone are not sufficient for AMOUNT. A numeric string used as a code, identifier, model name, or shared value is not labeled unless it expresses a measurable or countable quantity.

For TIMEDATE durations and ages, the time unit determines the class and stays inside the span (15 minutes, 19-year-old), while external modifiers such as about and up to stay outside.

• Record 33: about $ <AMOUNT>784,700.78</AMOUNT>.
• Record 85: <AMOUNT>6</AMOUNT>ft and <TIMEDATE>two months ago</TIMEDATE>.
• Record 9: less than <TIMEDATE>six hours</TIMEDATE> a night.
• Record 29: 3339 and 483 are numeric identifiers, not quantities.

Reason: this applies the baseline minimal-span principle consistently and prevents identifiers from being learned as quantities.

A3. Person names, nicknames, and usernames

Adjacent given and family names referring to one person form one PERSON span. Titles and possessive markers stay outside. A nickname inside a full name may remain inside the same span, including its quotation marks. A username or explicitly introduced nickname is labeled only when the context presents it as a real person reference. Structural dialogue placeholders are not labeled.

• Record 1: <PERSON>Manny Pacquiao</PERSON>, not two separate spans.
• Record 73: <PERSON>Norville "Shaggy" Rogers</PERSON>.
• Record 91: Person1 and Person2 are structural placeholders.

Reason: this extends the baseline multi-word name rule consistently to nicknames and usernames.

A4. Product, organization, and published work

A named application, game, platform, medicine, font, device, or branded good is PRODUCT when the context refers to the artifact or service. The company that operates or manufactures it is ORGANIZATION. A film, song, book, article, report, or other publication title is WORKOFART.

• Record 13: Discord is PRODUCT in the context on Discord.
• Record 26: Garamond is a font and is PRODUCT.
• Record 99: Project Performance Audit Report is WORKOFART.

Reason: this resolves recurring semantic confusion while following the baseline context rule.

A5. Occupations versus departments and activities

JOB is used only for an occupational title or formal work role in the sentence. A department, team, sector, field, activity, or generic organizational function is not JOB. A specifically named department may be ORGANIZATION.

• Record 66: Business Intelligence Director is JOB; market analysis division and analytics team are unlabeled.
• Record 77: Geriatric Medicine Department is ORGANIZATION, not JOB.
• Record 80: psychology and military are not occupations.

Reason: this prevents topics and organizational units from being learned as occupations.

A6. Locations, establishments, and companies

Named companies, sports teams, restaurants, casinos, and establishments remain ORGANIZATION even when someone visits them. Geographic areas, streets, addresses, landmarks, and physical facilities are LOCATION when the context emphasizes the place. The same surface form may therefore receive different labels in different contexts.

• Record 65: Joyride Taco House is a restaurant and is ORGANIZATION.
• Record 69: McDonald's is ORGANIZATION when referring to the company, but may be LOCATION when the text explicitly refers to being inside a physical restaurant.
• Record 56: West Wing Studio is LOCATION when described as a physical facility.

Reason: this makes the baseline context rule explicit for establishment/place confusion.

A7. No nested labels inside work titles

The dataset uses flat, non-overlapping spans. If another entity name appears inside a complete work title, the whole title is labeled WORKOFART and no nested span is created.

• Record 87: <WORKOFART>Diversity Rates in Mercer County, West Virginia Evaluation Essay</WORKOFART> is one span.

Reason: this keeps annotations compatible with flat BIO token classification and removes conflicting overlaps.

A8. Proposed names

An expression explicitly proposed as a name is labeled according to the intended entity type even if the object has not yet been released.

• Record 2: proposed publication titles are WORKOFART.
• Record 96: proposed pet-collar product names are PRODUCT.

Reason: the text explicitly presents these expressions as names; release status does not change the naming context.

A9. Record-removal quality threshold

A record is removed only when severe OCR/encoding corruption, an unrelated nonsensical word list, or truncation makes reliable annotation impossible. Ordinary spelling and grammar errors are retained as real language variation.

• Record 61: severe OCR corruption prevents reliable span selection.
• Record 80: keyword spam and nonsensical lists dominate the text.
• Record 92: generated nonsense and mixed-language fragments dominate the record.

Reason: removal is reserved for objective quality failures likely to harm the model, not used to avoid difficult annotation decisions.