fa-perdt-ner

Persian named-entity annotations over every sentence of the Persian Universal Dependency Treebank (UD_Persian-PerDT, PerUDT v1.0): 29,107 sentences, 494,163 tokens, in PerDT's own train/dev/test split, on the tokenization a spaCy --merge-subtokens conversion of the treebank produces.

Two configurations, same sentences, same tokens, different labels:

config	labels	entities	what it is
silver (default)	PER LOC ORG DAT MON TIM PCT	15,807	PerDT's own NER layer, realigned onto the UD tokenization. Produced by the PerDT authors with the Beheshti-NER tagger plus manual corrections; silver, not gold.
llm	PER LOC ORG DAT	15,144	A fresh annotation of the same sentences by a large language model, against a written guideline (v2.2) inferred from the silver layer. Four labels only.

Read this before comparing anything. llm has four labels; silver has seven. A model trained on one and scored on the other must be restricted to the four shared types, or every correct MON/TIM/PCT span becomes a false positive. And neither configuration is human gold: every agreement figure on this card is agreement between two annotators, not accuracy.

The text and split are PerDT's, CC BY-SA 4.0; this dataset is Adapted Material under the same licence. See Licence and attribution.

Loading

from datasets import load_dataset

silver = load_dataset("Phazel/fa-perdt-ner", "silver")   # 7 labels
llm = load_dataset("Phazel/fa-perdt-ner", "llm")         # 4 labels

row = llm["test"][0]
row["id"]        # PerDT sent_id, e.g. "test-s1"
row["tokens"]    # ["این", "موافقت‌نامه", "را", ...]
row["ner_tags"]  # ["O", "O", "O", ...]  IOB2 strings, parallel to tokens
row["entities"]  # [{"start": 3, "end": 5, "label": "ORG"}, ...]  token offsets, end exclusive

Splits are named train, validation (PerDT dev) and test. ner_tags are IOB2 strings, not a ClassLabel, so the two configurations can share one loader; map them yourself if a trainer wants integers.

Both configurations also ship as two-column IOB2 (<config>/{train,dev,test}.iob, tab separated, blank line between sentences), which spacy convert reads directly:

python -m spacy convert llm/train.iob corpus/ --converter ner --lang fa

The tokenization is exactly what spacy convert --merge-subtokens yields from the PerDT CoNLL-U files, so these files can be paired with a parser or tagger trained on the treebank without re-alignment. That is how the fa_core_news_* pipelines were built.

Fields

silver/*.jsonl:

field	type
id	string	PerDT sent_id (train-s1, dev-s12, ...)
tokens	list[string]	merged-subtoken surface forms, in order
ner_tags	list[string]	IOB2 tag per token

llm/*.jsonl adds:

field	type
entities	list[{start, end, label}]	token offsets, end exclusive; the spans ner_tags encodes
prompt_hash	string	3c5f435761a2 for every row: hash of the guideline+prompt that produced it
error	string or null	"missing" on the 1 sentence(s) the model refused to annotate: train-s16861

Statistics

silver:

split	sentences	tokens	entities	PER	LOC	ORG	DAT	MON	TIM	PCT
train	26196	445587	14228	4847	4954	2643	1323	205	135	121
dev	1456	24751	773	255	260	151	79	11	8	9
test	1455	23825	806	297	273	144	69	10	9	4

llm:

split	sentences	tokens	entities	PER	LOC	ORG	DAT
train	26196	445587	13644	4771	5011	2278	1584
dev	1456	24751	731	251	258	135	87
test	1455	23825	769	281	271	133	84

Entity counts are spans, from the IOB2 tags. Token counts are the same in both configurations by construction.

Provenance

Text and split

PerDT (Rasooli et al. 2013) is ~29k sentences of contemporary Persian across news, academic, magazine and fiction genres. Its UD conversion (Rasooli et al. 2022) is what was downloaded, at the URLs and MD5 checksums recorded in manifest.json. The 89/5/5 split is PerDT's own; no sentence was moved, dropped or deduplicated.

Tokenization: the CoNLL-U multiword tokens (pronominal clitics, enclitic copula) are merged back into their surface form, i.e. spacy convert --merge-subtokens. پدرم is one token, not پدر + م. This is the tokenization a spaCy fa tokenizer can actually produce, and the one every Phazel/fa_* pipeline is trained on.

silver

PerDT ships an entity layer at not-to-release/Dadegan with NER tag/, two-column IOB2 on the original Dadegan tokenization, which matches the released UD tokenization in only 57-62% of sentences (the NER files drop some copulas and auxiliaries). The PerDT README states this layer was produced with the BERT-based Beheshti-NER tagger (Taher, Hoseini & Shamsfard 2020) "with manual corrections to extend recall", and was used to mark PROPN in the treebank.

scripts/transfer_perdt_ner.py (in the pipeline repo) aligns each sentence's two token sequences with difflib and moves a span only if every one of its tokens maps and the result stays contiguous. Anything else is dropped rather than guessed. Transfer rate: 99.86% of spans on train, 99.74% on dev, 99.51% on test.

Known artefacts inherited from that layer, left as they are: the opening ( of a parenthetical blessing formula is tagged I-PER a few hundred times; a handful of entities carry inconsistent boundaries for the same string. The llm guideline documents the ones it found (annotation/GUIDELINES.md).

llm

A second annotation of the same sentences, produced in September 2026 by a large language model prompted with annotation/prompt-v2.2.md, which embeds annotation/GUIDELINES.md v2.2 in full. Pipeline, all in annotation/scripts/: select_sentences.py → build_requests.py (batches of 50 sentences, one guideline copy per batch) → annotate.js (calls the model, verifies the returned id set, retries once, then bisects a failing batch) → validate_responses.py (re-aligns each returned entity string onto the token sequence, rejects anything that does not align to whole tokens) → export_iob.py.

The guideline was inferred from the silver layer, not written from linguistic principles: each rule states the majority convention in the silver train split with the counts behind it (e.g. آقای outside a PER span 140/141 times), and calls the minority out as an error to correct. So the two configurations are not independent: llm is what the silver conventions look like when applied consistently.

Three entity types were deliberately dropped. MON, TIM and PCT are 3.2% of silver spans and the least consistent — the guideline could not find a majority convention to state for them. Consumers wanting those types should use silver.

Model provenance is incomplete and recorded honestly: annotate.js was run with the model alias default (the harness's session default) and the resolved model id, temperature and date were not written into the output rows. Re-running the same prompt on a sample agreed with the committed run at span F 0.95-0.98, so the committed files are the record of origin and cannot be reproduced bit-exactly. prompt_hash is stable, so a re-run against a changed guideline is detectable.

42 entities the model returned were rejected by validate_responses.py and are listed in llm/dropped-entities.jsonl with the sentence they came from: 36 no-align, 5 bad-label, 1 overlap. no-align means the returned string does not cover whole tokens, mostly a name the model separated from its clitic (فاطمه against the token فاطمه‌ام) or a span the model re-spelled; bad-label is a label outside the four; overlap is a span inside another.

Agreement between the two configurations

Exact span match, silver restricted to the four shared labels as the reference, from llm/agreement.json:

split	PER F	LOC F	ORG F	DAT F	micro P	micro R	micro F
train	0.716	0.816	0.675	0.692	0.746	0.739	0.742
dev	0.723	0.795	0.706	0.747	0.755	0.741	0.748
test	0.734	0.827	0.679	0.706	0.761	0.747	0.754

Adjudicating every test-split disagreement span by span (LLM-judged, so an estimate) put the llm annotations at 0.93-0.95 precision against silver's 0.84: silver is wrong or mis-bounded on roughly 126 of its 783 test spans. That number is why llm exists; it is not a measurement.

Training the same model on each

Same spaCy NER architecture, same hyperparameters, same sentences; micro F over PER/LOC/ORG/DAT, out-of-scope entities removed from both prediction and reference:

trained on \ scored on	silver test	llm test
silver	71.98	67.27
llm	67.59	79.94

The diagonal says the llm labels are more internally consistent (+7.96 F for an identical model). The near-symmetric off-diagonal says the two conventions are genuinely different, not one a subset of the other. ORG recall does not improve under llm (68.1 → 66.9): the head-noun-plus-specifier construction (کمیسیون آموزش) is under-tagged in both. Details in the pipeline repo's docs/MODELS.md §10.

Limitations

• No human gold. Nothing on this card measures accuracy. A 200-sentence blind human sample over the test split is drawn and unfilled; until it is, prefer silver for reporting numbers comparable to prior PerDT work, and treat llm as a consistency-improved training set whose real recall is unknown.
• llm has four labels. A seven-label model evaluated on it, or vice versa, needs the label restriction described above.
• Silver is silver. Tagger output with corrections of unknown extent; the PerDT authors used it to mark PROPN, not as an NER benchmark.
• One sentence in llm is unannotated (train-s16861, a content-filter refusal); its ner_tags are all O and error is "missing".
• Genre. Contemporary written Persian, news-heavy. No social media, no dialect.
• Tokenization is spaCy-shaped. Clitics are attached. Systems that split them need to re-align.

Licence and attribution

CC BY-SA 4.0, inherited: PerDT is CC BY-SA 4.0 and this is Adapted Material, so ShareAlike applies to both configurations and to anything trained on them that constitutes an adaptation. Full text and notice in LICENSE.

Please credit the treebank:

@inproceedings{rasooli-etal-2022-persian,
  title     = {The {P}ersian Dependency Treebank Made Universal},
  author    = {Rasooli, Mohammad Sadegh and Safari, Pegah and Moloodi, Amirsaeid and Nourian, Alireza},
  booktitle = {Proceedings of the Thirteenth Language Resources and Evaluation Conference},
  year      = {2022},
  pages     = {7078--7087},
  url       = {https://aclanthology.org/2022.lrec-1.766}
}

@inproceedings{rasooli-etal-2013-development,
  title     = {Development of a {P}ersian Syntactic Dependency Treebank},
  author    = {Rasooli, Mohammad Sadegh and Kouhestani, Manouchehr and Moloodi, Amirsaeid},
  booktitle = {Proceedings of NAACL-HLT 2013},
  year      = {2013},
  url       = {https://aclanthology.org/N13-1031}
}

and, for the silver layer's tagger:

@article{taher-etal-2020-beheshti,
  title   = {Beheshti-{NER}: {P}ersian named entity recognition Using {BERT}},
  author  = {Taher, Ehsan and Hoseini, Seyed Abbas and Shamsfard, Mehrnoush},
  journal = {arXiv preprint arXiv:2003.08875},
  year    = {2020}
}

If you use the llm configuration or the annotation kit, cite this dataset:

@misc{fazeli-2026-fa-perdt-ner,
  title  = {fa-perdt-ner: {P}ersian NER over the {PerDT} treebank, silver and {LLM}-relabelled},
  author = {Fazeli, Kiyarash},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Phazel/fa-perdt-ner}
}

Source, build script and the full annotation history: https://github.com/Fazel94/spacy-persian (annotation/, scripts/annotation/). manifest.json in this repo records the commit each file was built from.