wiki-gec-distill-eu — 290,935 Wikipedia sentences labelled by an 8B GEC teacher

Each row is a Basque Wikipedia sentence plus the output of the teacher model itzune/latxa-8b-gec-lora-v2, with the teacher's sequence log-probabilities for the corrected and the unchanged sentence. Those two numbers let you choose your own precision/recall trade-off after the fact — no re-running the model.

This is the corpus that produced the 36,878 extra training pairs behind itzune/gector-eus-v5: its real Wikipedia F0.5 went from 25.2 (v4.1) to 44.1.

Composition

pool	rows	what it is
pre-edit	195,747	pre-revision sentences from mined Wikipedia pairs — a human later edited them, so error density is high
stable	95,188	sentences that survived many revisions — mostly clean text
total	290,935	52,626 characters on average after filtering to 25–400 chars

The teacher proposes an edit in 25.8% of the rows (34% of pre-edit, 8.6% of stable).

Fields

field	description
id	pre-edit-* / stable-*
pool	pre-edit or stable
title	Wikipedia article title (attribution)
source	input sentence
teacher_out	teacher's greedy correction (equal to source when it proposes nothing)
lp_out	teacher-forced log P(teacher_out + eos │ prompt)
lp_copy	teacher-forced log P(source + eos │ prompt)
margin	lp_out − lp_copy (positive = teacher prefers its own correction)
edited	teacher_out != source
rev, rev_parent, rev_ts, rev_user, rev_comment	revision provenance (present when the sentence came from a mined revision pair)
rev_n_revisions	revision count for stable sentences

Using it: the copy-vs-edit margin

from datasets import load_dataset

ds = load_dataset("itzune/wiki-gec-distill-eu", split="train")

def pairs(margin=-1.0):
    """Teacher-edit training pairs, filtered by confidence (v5 used -1)."""
    return [
        {"source": r["source"], "target": r["teacher_out"]}
        for r in ds
        if r["edited"] and r["margin"] > margin
    ]

threshold	edits kept
margin > -1 (v5 recipe, before guards)	69,611
margin > 0	61,738
margin > 1	49,592

The GECToR v5 build additionally drops edits that touch capitalised names (Felipe → Filipe) or change sentence length, giving the final 36,878 pairs, and samples 50% of the unedited sentences as clean (no-op) records.

How it was produced

1. Mine sentence-level revision pairs from the Basque Wikipedia dump euwiki-20261101 (scripts/mine_wiki_revisions.py).
2. Build a labelling pool of pre-edit + stable sentences, 25–400 characters, deduplicated, excluding every article that contributed to the evaluation sets and any sentence present in data/real_eval or the synthetic blind slice (scripts/build_distill_corpus.py).
3. Label with itzune/latxa-8b-gec-lora-v2 under vLLM: greedy correction, then two teacher-forced scoring passes (scripts/llm_label_vllm.py). ~17 sentences/s on one L40.
4. Turn labels into GECToR raw data with the margin and name/size guards (scripts/build_distill_data.py).

Limitations

• Teacher output, not ground truth. The teacher over-corrects on clean text; unfiltered, its edits hit ~8% of clean sentences. Always apply a margin.
• Domain match: sentences are Wikipedia prose, the same genre as most Basque GEC test material — gains measured on Wikipedia test sets are partly in-domain. Evaluation articles were excluded, and no sentence overlaps the eval sets.
• stable sentences are probably clean, not verified: some contain typos that survived many revisions.
• lp_out/lp_copy are teacher-forced log-probs over the whole sentence, computed with the same prompt; they are only comparable within a row.
• Sentence-level only: no document context, and the pool is capped at 400 characters to keep labelling throughput high.

Licensing and attribution

Source text is the Basque Wikipedia dump euwiki-20261101, CC-BY-SA-4.0 — title and revision metadata are included for attribution. The corrections and log-probabilities are outputs of a model derived from HiTZ/Latxa-Llama-3.1-8B-Instruct, distributed under the Llama 3.1 Community Licence; that licence restricts using outputs to train other large language models. Review those terms before commercial use — the intended consumer here is a 125M token-classification model.

Citation

@misc{ezpeleta2026wikigecdistill,
  author       = {Xabi Ezpeleta},
  title        = {wiki-gec-distill-eu: Basque Wikipedia sentences labelled by an
                  8B grammatical-error-correction teacher},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/itzune/wiki-gec-distill-eu}},
  note         = {Teacher: \url{https://huggingface.co/itzune/latxa-8b-gec-lora-v2}}
}

References

• Teacher model: itzune/latxa-8b-gec-lora-v2
• Student model: itzune/gector-eus-v5
• Human-edited corpus: itzune/horkonpon-corpus
• Basque Wikipedia dump euwiki-20261101, dumps.wikimedia.org/euwiki (CC-BY-SA-4.0)
• Etxaniz, J. et al. (2024). Latxa: An Open Language Model and Evaluation Suite for Basque. arXiv:2403.20266
• Omelianchuk, K. et al. (2020). GECToR — Tag, Not Rewrite. BEA@ACL. arXiv:2005.12592
• vLLM — labelling throughput
• Pipeline code: itzune/gector-eus-v2 (scripts/build_distill_corpus.py, scripts/llm_label_vllm.py, scripts/build_distill_data.py)