Persian (Farsi) pipelines for spaCy

Trained spaCy pipelines for Persian, installable with pip. spaCy has never shipped an official one, and spacy.blank("fa") only gives you a tokenizer and stop words. Ten packages across four tiers: sm (hash embeddings), md and lg (floret static vectors), trf (fine-tuned ParsBERT). Each of sm/md/lg ships as syntax only (fa_dep_news_*), syntax + NER (fa_core_news_*), or NER alone (fa_ent_news_*); trf ships as fa_core_news_trf only.

pip install https://huggingface.co/Phazel/fa_core_news_sm/resolve/main/fa_core_news_sm-3.8.0-py3-none-any.whl

>>> import spacy
>>> nlp = spacy.load("fa_core_news_sm")

>>> doc = nlp("محمدرضا شجریان در مشهد به دنیا آمد.")
>>> [(t.text, t.pos_, t.lemma_, t.dep_) for t in doc][:2]
[('محمدرضا', 'PROPN', 'محمدرضا', 'nsubj'), ('شجریان', 'PROPN', 'شجریان', 'flat:name')]
>>> doc.ents
(محمدرضا شجریان, مشهد)

>>> doc = nlp("وزارت نفت روز شنبه قیمت بنزین را ۱۰ درصد افزایش داد.")
>>> [(e.text, e.label_) for e in doc.ents]
[('وزارت نفت', 'ORG'), ('روز شنبه', 'DAT'), ('۱۰ درصد', 'PCT')]

Results

Compared against Hazm (https://github.com/roshan-research/hazm) and en_core_web_sm (English reference).

Metric	spacy-persian fa_core_news_trf	Hazm (Persian toolkit)	en_core_web_sm (English reference)
POS accuracy	97.63% UPOS	98.8% own tagset¹	97.29% PTB XPOS²
Lemma accuracy	97.31%	89.9%¹	not reported²
Dependency UAS / LAS	93.87% / 90.79%	92.30% / 89.15%¹	91.77% / 89.92%²
NER F-score	82.89%	not reported¹	84.33%²

¹ Hazm's own README, https://github.com/roshan-research/hazm#evaluation: POSTagger 98.8% on Hazm's EZ-augmented tagset, which is not UPOS; Lemmatizer 89.9%; SpacyDependencyParser UAS 92.30 / LAS 89.15. It reports no NER score.

² en_core_web_sm 3.8.0 meta.json, https://github.com/explosion/spacy-models/blob/master/meta/en_core_web_sm-3.8.0.json: tag_acc 0.9729, dep_uas 0.9177, dep_las 0.8992, ents_f 0.8433. That pipeline reports no pos_acc and no lemma_acc, because OntoNotes has no UPOS or lemma layer.

The three columns were scored on different test data.

Packages

Package	Components	Licence	Score	Wheel
fa_dep_news_sm	tok2vec, tagger, morphologizer, trainable_lemmatizer, parser	CC BY-SA 4.0	LEMMA 97.91	7.9 MB
fa_core_news_sm	the above plus ner	CC BY-SA 4.0	ENTS_F 71.87	13.5 MB
fa_ent_news_sm	ner alone (own embedded tok2vec)	CC BY-SA 4.0	ENTS_F 71.87	5.9 MB
fa_dep_news_md	same as fa_dep_news_sm, plus floret vectors	CC BY-SA 4.0	LEMMA 97.96	62.6 MB
fa_core_news_md	same as fa_core_news_sm, plus floret vectors	CC BY-SA 4.0	ENTS_F 74.71	68.5 MB
fa_ent_news_md	ner alone (own embedded tok2vec), plus floret vectors	CC BY-SA 4.0	ENTS_F 74.71	60.6 MB
fa_dep_news_lg	same as fa_dep_news_sm, plus full-wiki floret vectors	CC BY-SA 4.0	LEMMA 98.08	229.3 MB
fa_core_news_lg	same as fa_core_news_sm, plus full-wiki floret vectors	CC BY-SA 4.0	ENTS_F 75.94	235.2 MB
fa_ent_news_lg	ner alone (own embedded tok2vec), plus full-wiki floret vectors	CC BY-SA 4.0	ENTS_F 75.94	227.3 MB
fa_core_news_trf	transformer, tagger, morphologizer, trainable_lemmatizer, parser, ner	encoder states no licence, §8	ENTS_F 82.89, LAS 90.79	608.2 MB

These scores are from spacy benchmark accuracy and ship in each package's meta.json under performance.

Raw fa.floret and fa.vec exports of the lg tier's 200k-row table are in fa-floret-wiki-vectors.

Tier comparison

The md tier adds a 50k x 300d floret vector table trained on the first 400,000 Persian Wikipedia articles. Its config differs from sm by exactly one line (include_static_vectors), so the columns below isolate what the vectors buy. Full breakdown in docs/MODELS.md §6.

Metric	sm	md	lg	trf	Reference
TOKEN_ACC / TOKEN_F	99.96 / 99.11	99.96 / 99.11	99.96 / 99.11	99.96 / 99.11
TAG_ACC (XPOS)	95.96	96.25	96.55	97.62
POS_ACC (UPOS)	96.24	96.64	96.68	97.63
MORPH_ACC	96.29	96.64	96.70	97.82
LEMMA_ACC	97.91	97.96	98.08	97.31
SENTS_F	99.25	99.28	99.18	97.35
DEP_UAS	89.69	90.52	90.96	93.87	hazm-bert-dependency-parser: 92.46
DEP_LAS	85.15	86.34	86.60	90.79	hazm-bert-dependency-parser: 89.34
ENTS_P	77.67	76.56	81.51	84.06
ENTS_R	66.87	72.95	71.09	81.76
ENTS_F	71.87	74.71	75.94	82.89
Speed (940MX, batch 32)	10,235 words/s	9,058 words/s	9,215 words/s	1,106 words/s
Wheel size	13.5 MB	68.5 MB	235.2 MB	608.2 MB

Reference cells are hazm-bert-dependency-parser's own meta.json, trained on the same treebank (docs/MODELS.md §4).

trf leads on every metric except lemmatization and sentence segmentation. It is also the only tier to clear the hazm-bert-dependency-parser DEP_LAS reference of 89.34. It needs a GPU in production (187 words/s on the laptop CPU), and its HooshvareLab/bert-base-parsbert-uncased encoder states no licence, so the published wheel carries a redistribution warning in its meta.json and its terms are unknown (docs/MODELS.md §8).

Entity scores are fa_core_news_* on the PerDT NER test split; per-label breakdown and caveats are in Named entity recognition.

sm, md and lg were trained on a 4-core i5-7200U, CPU only: sm took 1h27m for syntax plus 17 min for NER, md 1h54m plus 25 min (the two md runs overlapped, so wall clock overstates each), lg about 2h08m plus 13 min. trf took 1h58m on a rented Colab T4.

Vector packages

Standalone floret vector packages (vectors only, pipeline: []), usable as --paths.vectors for your own training or as a plain embedding table:

# 50k rows x 300d, first 400,000 Persian Wikipedia articles (the md tier's table)
pip install https://huggingface.co/Phazel/fa_floret_400k/resolve/main/fa_floret_400k-0.1.0-py3-none-any.whl
# 50k rows x 300d, full Persian Wikipedia dump
pip install https://huggingface.co/Phazel/fa_floret_full_wiki/resolve/main/fa_floret_full_wiki-0.1.0-py3-none-any.whl
# 200k rows x 300d, full Persian Wikipedia dump, 5 epochs (the lg tier's table)
pip install https://huggingface.co/Phazel/fa-floret-wiki-vectors/resolve/main/fa_floret_wiki_200k-0.1.0-py3-none-any.whl

Throughput

Median of repeated nlp.pipe passes over the 146-document PerDT test split (23,825 tokens), timing the pipe only, warmup discarded. Reproduce with python scripts/benchmark_throughput.py <model> --gpu-id <n>, which writes the raw records to metrics/throughput-*.json.

Tier	CPU, i5-7200U	GPU, GeForce 940MX	CPU, Xeon @ 2.00GHz	GPU, Tesla T4
sm	5,484	10,235
md	5,408	9,058
lg	4,715	9,215
trf	187	1,106	336	8,320

trf is 29x slower than sm on the same CPU. The Xeon and T4 columns come from one Colab VM, a 25x GPU speedup. The CPU tiers sit within 15% of each other, so the bottleneck is the parser and lemmatizer, not the tok2vec lookup. Laptop spread is about 10% with thermal state. Running trf on the 940MX needs a cu126 torch build, see docs/MODELS.md §9.

Against other Persian toolkits

The same 146 documents as raw text, full pipeline, same laptop CPU. Reproduce with scripts/benchmark_toolkit.py.

Toolkit	Components	Words/s
fa_dep_news_sm	tokenizer, tagger, morphologizer, lemmatizer, parser	13,073
fa_core_news_sm	the same, plus NER	8,831
UDPipe 1.4, persian-seraji-ud-2.5 model	tokenizer, tagger, lemmatizer, parser	1,929
DadmaTools 2.3.6	tokenizer, lemmatizer, POS tagger, parser, on XLM-RoBERTa	96
Stanza 1.14, perdt models	tokenizer, MWT, POS tagger, lemmatizer, parser	82
Hazm 0.12.1	normalizer, tokenizers, POS tagger, lemmatizer, MaltParser	25

fa_dep_news_sm is 7x faster than UDPipe, 136x faster than DadmaTools and 160x faster than Stanza. md and lg stay within 15% of sm in the table above, so they lead too; trf, at 187 words/s, does not. DadmaTools and Stanza ran on the first 16 documents (2,396 words), the others on all 146. Every row except Hazm was measured on 2026-09-28 with one thread per process; Hazm was timed on 2026-08-11 in one 933-second pass. The table above comes from another day with different background load, which is why fa_core_news_sm reads 5,484 there and 8,831 here: compare within a table, not across. Speed only: the UDPipe 1 models (UD 2.5) predate PerDT, so its row uses Seraji.

Named entity recognition

Seven labels: LOC, PER, ORG, DAT, MON, TIM, PCT. They come from PerDT's own not-to-release/Dadegan with NER tag/ layer, transferred onto this pipeline's tokenization by difflib at a 99.86% alignment rate (scripts/transfer_perdt_ner.py). Spans that could not be aligned exactly were dropped rather than guessed. That layer is silver: PerDT's README states it was produced by the BERT-based Beheshti-NER tagger with manual corrections for recall, so the ENTS_F numbers below partly reflect agreement with that tagger, not with human annotation.

Both that realigned layer and a four-label LLM relabelling of the same sentences (annotation/, guideline v2.2) are published as Phazel/fa-perdt-ner, CC BY-SA 4.0, keyed by PerDT sent_id; spacy project run hub-dataset rebuilds it. The relabelling is measured against the silver layer in docs/MODELS.md §10 and ships no model yet.

ner runs standalone with its own embedded tok2vec (fa_ent_news_sm, fa_ent_news_md, fa_ent_news_lg), or bundled into fa_core_news_sm, fa_core_news_md and fa_core_news_lg alongside the syntax pipeline. In trf it is trained jointly against the shared transformer instead, so there is no standalone trf variant.

Label	sm F	md F	lg F	trf F	Train examples
LOC	80.24	84.05	83.66	87.78	4,954
PER	65.29	68.18	72.63	81.88	4,847
ORG	68.77	70.25	71.01	78.50	2,643
DAT	74.45	76.19	70.83	82.52	1,323
MON	73.68	84.21	88.89	88.89	205
TIM	66.67	66.67	61.54	50.00	135
PCT	57.14	33.33	57.14	33.33	121

MON, TIM and PCT have single-digit support in the test split, so their deltas are one or two entities changing hands, not signal. PER, LOC and ORG carry the split. The md gain over sm (ENTS_F 71.87 to 74.71) is almost entirely recall (+6.08), the lexical prior static vectors give rare proper nouns that hash embeddings never had. trf adds another +6.95 F over lg, again mostly recall (71.09 to 81.76), and its largest per-label gains are PER (+9.25) and DAT (+11.69).

Install

# syntax + NER, 13.5 MB
pip install https://huggingface.co/Phazel/fa_core_news_sm/resolve/main/fa_core_news_sm-3.8.0-py3-none-any.whl
# syntax only, 7.9 MB
pip install https://huggingface.co/Phazel/fa_dep_news_sm/resolve/main/fa_dep_news_sm-3.8.0-py3-none-any.whl
# NER only, 5.9 MB
pip install https://huggingface.co/Phazel/fa_ent_news_sm/resolve/main/fa_ent_news_sm-3.8.0-py3-none-any.whl

For the vector tiers, swap sm for md or lg in both the repo name and the filename:

pip install https://huggingface.co/Phazel/fa_core_news_md/resolve/main/fa_core_news_md-3.8.0-py3-none-any.whl
pip install https://huggingface.co/Phazel/fa_core_news_lg/resolve/main/fa_core_news_lg-3.8.0-py3-none-any.whl

fa_core_news_trf is versioned 1.0.0, not 3.8.0, and needs spacy-transformers:

pip install spacy-transformers
pip install https://huggingface.co/Phazel/fa_core_news_trf/resolve/main/fa_core_news_trf-1.0.0-py3-none-any.whl

Caveats

• Some lemmas contain a space. Multiword tokens were merged, so کتاب‌هایش is one token tagged N_IANM_PR_JOPER with lemma کتاب او. This affects about 1.5% of tokens.
• doc.noun_chunks under-fires. spacy/lang/fa/syntax_iterators.py upstream matches ClearNLP labels that do not exist in Universal Dependencies. Bug analysis and proposed upstream patch in docs/upstream/fa-noun-chunks.md.

Build

The corpora download as checksummed assets; the md and lg tiers also need a floret wheel, described below. Python 3.12:

python -m venv .venv
.venv/bin/python -m pip install -U pip
.venv/bin/python -m pip install -r requirements.txt

.venv/bin/python -m spacy project assets      # download + checksum the corpora
.venv/bin/python -m spacy project run all     # -> fa_dep_news_sm + fa_core_news_sm
.venv/bin/python -m spacy project run ent     # -> fa_ent_news_sm, NER alone

Command	What it does
inspect	annotation coverage of the treebanks (scripts/inspect_treebanks.py)
convert-ud	CoNLL-U to DocBin with --merge-subtokens, plus the tokenizer-agreement report
transfer-ner	align PerDT's NER layer onto that tokenization by difflib (scripts/transfer_perdt_ner.py)
convert-ner	transferred IOB2 to DocBin
debug-data, debug-data-ner	spacy debug data on both corpora before spending CPU
train-dep	tagger + morphologizer + trainable_lemmatizer + parser
train-ner	the ner component, with its own embedded tok2vec
finalize-dep	write fa_dep_news_sm metadata: sources, licence, notes (scripts/finalize_pipeline.py)
evaluate-dep	spacy benchmark accuracy on the held-out UD test split
assemble-core	source ner into the dep pipeline to produce fa_core_news_sm
evaluate-core	score the assembled pipeline on both test splits
finalize-meta	re-run finalize on both, folding test scores into meta.json["performance"]
package	build wheels + sdists for both
smoke	run both pipelines over Persian text and print every annotation layer
finalize-ent	write fa_ent_news_sm metadata from the standalone NER run
evaluate-ent	spacy benchmark accuracy for the NER-only package
package-ent	build the fa_ent_news_sm wheel + sdist

The two training runs are single-threaded and independent, so they can run concurrently.

The vector and transformer tiers are separate workflows:

.venv/bin/python -m spacy project run md      # -> fa_dep_news_md, fa_core_news_md
.venv/bin/python -m spacy project run lg      # -> fa_dep_news_lg, fa_core_news_lg
.venv/bin/python -m spacy project run trf     # -> fa_core_news_trf, GPU only

The md and lg workflows stop at the dep and core packages. fa_ent_news_lg is built by finalize-ent-lg, evaluate-ent-lg and package-ent-lg, which no workflow calls; run them by name.

md and lg start by unpacking a floret vector wheel that spacy project assets does not download, because it is built by this project rather than fetched. Put it in the repo root under the exact filename project.yml expects (vars.floret_wheel, vars.floret_lg_wheel):

curl -L -o fa_floret-0.1.0-py3-none-any-400k-documents.whl \
  https://huggingface.co/Phazel/fa_floret_400k/resolve/main/fa_floret_400k-0.1.0-py3-none-any.whl
curl -L -o fa_floret-0.1.0-py3-none-any-full-wiki-200k-5epoch.whl \
  https://huggingface.co/Phazel/fa-floret-wiki-vectors/resolve/main/fa_floret_wiki_200k-0.1.0-py3-none-any.whl

trf additionally needs spacy-transformers and a real GPU (vars.gpu_trf is 0); the 940MX needs a cu126 torch build in a separate venv, see docs/MODELS.md §9.

Design decisions

1. --merge-subtokens. spaCy has no multiword-token layer, and PerDT splits pronominal clitics (پدرم into پدر + م). Measured on dev, merging gives token F 0.9887 against 0.9823 for the split version, costing 34 composite XPOS tags on 1.5% of tokens. Without it, 1.5% of gold tokens are boundaries the shipped tokenizer can never produce. See scripts/tokenization_report.py.
2. ner carries its own tok2vec. A Tok2VecListener only resolves inside the pipeline it was trained in, so a listener-based component cannot be sourced elsewhere. configs/fa_ner_sm.cfg embeds the tok2vec instead, as en_core_web_sm does.
3. morphologizer + trainable_lemmatizer instead of attribute_ruler + rule lemmatizer. The English pipelines derive UPOS from PTB tags by rule because OntoNotes has no UPOS. UD gives gold UPOS, FEATS and lemmas, which yields real pos_acc, morph_acc and lemma_acc numbers instead of unmeasurable rule coverage.
4. PerDT, not Seraji: 3.7x more tokens, and Seraji has no PROPN tag.

Why not Hazm's own models

Hazm publishes spaCy-format pipelines on the HF Hub, so it was the obvious starting point. Four problems:

• Its trainable models are pycrfsuite CRFs (hazm/sequence_tagger.py). The repo contains no config.cfg and no spacy train; the Spacy* classes only download pretrained pipelines.
• Those pipelines are three single-task models (transformer + tagger, transformer + parser, transformer + chunker), each version: 0.0.0 with an empty license field, pinned to spaCy 3.6. Using all three costs three ParsBERT forward passes and gives no shared Doc.
• Its tokenizer is incompatible with UD tokenization: the normaliser fuses ZWNJ affixes and join_verb_parts() glues multi-word verb chains into single tokens.
• Most corpora it reads (Bijankhan, Peykare, Hamshahri, raw PerDT) sit behind peykaregan.ir or dadegan.ir under research-only terms.

It did confirm the corpus choice. Hazm's own spaCy parser was trained on modified_fa_perdt-ud-train.spacy, the same treebank used here.

More

• Pipeline inventory, corpus and licence analysis: docs/MODELS.md
• How spaCy models get published, and what upstream fa already has: docs/CONTRIBUTING-GUIDE.md
• The build: project.yml
• Language data comes from spacy/lang/fa upstream, whose stop word list came from Hazm.
• خلاصهٔ فارسی: README.fa.md