horkonpon-corpus — a Basque grammatical-error-correction corpus

Horkonpon (Basque, "hor konpon" ≈ "fix it there"). 248,008 training records of Basque sentence pairs with edit-level annotations (span, error category, severity, normative reference) for grammatical error correction, plus evaluation splits, mined human Wikipedia edits and two hand-labelled review sheets.

The corpus is the training source for itzune/gector-eus-v2, itzune/gector-eus-v5 and itzune/latxa-8b-gec-lora-v2, and the generator code lives at github.com/itzune/horkonpon-corpus.

Composition

config	split	records	what it is
v41	train	248,008	training pairs: 117,811 error · 95,121 clean · 35,076 suggestion
eval	standard	1,088	gold eval split (error + clean), document-disjoint from train
eval	multi	826	real multi-error slice (one more error injected, no comma deletions)
eval	morph_holdout	300	held-out Basque paradigm pairs (unseen verb/case forms)
wiki	edits	220,234	human edits mined from Basque Wikipedia revisions (2010–2026)
wiki	clean	119,210	sentences that survived many revisions (stable text)
review	errors	600	hand-labelled mined revision pairs (verdict: real error or noise)
review	clean	300	hand-labelled stable sentences (clean or not)

from datasets import load_dataset

train = load_dataset("itzune/horkonpon-corpus", "v41", split="train")
ev    = load_dataset("itzune/horkonpon-corpus", "eval", split="standard")
wiki  = load_dataset("itzune/horkonpon-corpus", "wiki", split="edits")
print(train[0])

Fields

v41, eval:

field	description
id	stable record id (hkp-*)
task	gec (or explain in the EXCGEC-format export)
nature	error (normative, EBE) · suggestion (editorial, Berria Estilo Liburua) · clean
source / target	erroneous / corrected sentence
edits	list of edit objects (see below)
categories_present	categories touched by the record
n_edits	number of edits (0 for clean records)
source_doc, source_domain, domain	carrier document id and register (news, encyclopedic, gazette, …)
license	licence of the carrier text (cc-by-sa, cc-by, public-domain)
carrier	carrier corpus (latxa-corpus-v2, llm-latxa-70b, mined-morphology)

Each entry of edits:

field	description
src_interval / tgt_interval	token span in source / target (token indices)
src_content / tgt_content	the strings involved
src_tokens / tgt_tokens	tokenised spans
category	spelling, morphology, punctuation, capitalization, zalantza, proper_noun, word_level, calque
error_type	fine-grained type (char_delete, paradigm_agreement, missing_semicolon, …)
error_severity	1–3
nature	error vs suggestion
ref	normative reference for the correction (EBE table, Berria style guide)
error_description	plain-Basque explanation

wiki/edits: title, rev, parent, ts, user, ip, bot, comment, source, target, changes (list of [old, new] strings).

wiki/clean: title, source, target, n_revisions.

review/*: id, shape (edit shape), source, target, change, comment, verdict (Y/N/…), fixed_target, notes.

How it was built

Two halves, both shipped here:

1. Injected errors (the v41 and eval configs). Carrier sentences come from professionally edited Basque prose filtered to commercially usable licences — news (latxa-corpus-v2), Basque government gazettes and encyclopedic text; NC-licensed sources (Addi, Ekaia) and custom-licence administrative sources are excluded at the stream level. Each clean carrier is pre-filtered with txukun's EBE-grounded rule engine: any sentence that trips a rule is discarded, never auto-corrected. 16 injectors (8 capitalisation/punctuation, 4 spelling, 3 morphology, 1 word-level) corrupt the clean side; sampling weights follow real error frequencies. Zalantza (EBE-attested doubtful words) and calque injectors use curated tables (803 pairs).
2. Mined human edits (the wiki config). Sentence-level revision pairs from the Basque Wikipedia dump euwiki-20261101, restricted to small edits, with the corrected sentence required to survive to the latest revision. The review config is the hand-labelled sample used to measure how much of this mined data is real error: spelling 63%, capitalisation 61%, punctuation 33%, multi-change 34%, inflection 24%, word insert 9%, swaps/deletes/rewrites 2–17%.

Quality gates: 100% structural pass, 0.59% oracle lints, 679 unit tests, document-level train/eval separation, and a language filter (lingua) that dropped 37,262 non-Basque records.

Models trained on it

model	notes	real-Wikipedia F0.5 (171 sentences)
itzune/gector-eus-v2	GECToR, v2 data	15.7
itzune/gector-eus-v5	GECToR, v4.1 + teacher distillation	43.7
itzune/latxa-8b-gec-lora-v2	LoRA on Latxa-8B	58.2

Limitations

• The v41 half is synthetic: errors are injected by rule, so their distribution is flatter and cleaner than real writing. Real-error shares are measured in review, not assumed.
• The wiki half is noisy: mined edits include vandalism, rewrites and stylistic changes — use review to calibrate any filter you build on it.
• Carrier sentences are professionally edited news/gazette prose, so models trained here underperform on social-media Basque.
• suggestion records are editorial preferences, not errors; keep them separate when evaluating error detection.
• Wikipedia-derived rows carry article titles and revision metadata for attribution, but some old revisions have hidden or anonymous authors.

Licensing and attribution

• v41, eval: CC-BY-SA-4.0. Carrier text is CC-BY / CC-BY-SA / public domain; each record carries its own license field and source_doc.
• wiki configs: derived from the Basque Wikipedia (euwiki-20261101) under CC-BY-SA-4.0; rev, rev_ts and user are included for attribution.
• EBE (Euskara Batuaren Eskuliburua) is used as a normative reference, not copied, so it imposes no additional licence.
• The raw 4.4 GB XML dump is not redistributed — fetch it from dumps.wikimedia.org/euwiki and re-mine with scripts/mine_wiki_revisions.py.

Citation

@misc{ezpeleta2026horkonpon,
  author       = {Xabi Ezpeleta},
  title        = {horkonpon-corpus: a Basque grammatical error correction corpus
                  with edit-level annotations},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/itzune/horkonpon-corpus}},
  note         = {CC-BY-SA-4.0. Code: \url{https://github.com/itzune/horkonpon-corpus}}
}

References

Normative sources

• Euskaltzaindia — Euskara Batuaren Eskuliburua (EBE), euskaltzaindia.eus
• Berria — Estilo Liburua, berria.eus/estiloliburua
• Beloki, Z. et al. — Basque verb-form / paradigm tables used for morphology injection

Data sources

• Basque Wikipedia dump euwiki-20261101, dumps.wikimedia.org/euwiki (CC-BY-SA-4.0)
• Latxa-Corpus-v2 (news carrier text, CC-BY-SA-4.0), HiTZ
• Basque government gazettes (BOPV/Euskadi.eus) — open data

Methods and related work

• Omelianchuk, K. et al. (2020). GECToR — Grammatical Error Correction: Tag, Not Rewrite. BEA@ACL. arXiv:2005.12592
• Ye, T. et al. (2025). EXCGEC: A Benchmark for Explainable Chinese Grammatical Error Correction. AAAI 2025 — annotation-schema inspiration.
• Bryant, C., Felice, M., Briscoe, T. (2017). Automatic Annotation and Evaluation of Error Types in Learner Corpora — M²-style span alignment used here.

Tooling

• txukun — EBE-grounded rule engine used as a clean-text oracle
• generator (GitHub) — corruption injectors, mining scripts, QA and tests
• lingua — language identification filter
• Hugging Face Datasets
• Model family: itzune/gector-eus-v2