Datasets:

itzune
/

eusgec-wiki-eval

---
license: cc-by-sa-4.0
language:
  - eu
pretty_name: eusgec-wiki-eval — hand-checked real Basque Wikipedia errors
size_categories:
  - n<1K
tags:
  - grammatical-error-correction
  - gec
  - basque
  - euskara
  - evaluation
  - wikipedia
configs:
  - config_name: default
    data_files:
      - split: errorful
        path: default/errorful-*.parquet
      - split: clean
        path: default/clean-*.parquet
  - config_name: review
    data_files:
      - split: judged
        path: review/judged-*.parquet
---
# eusgec-wiki-eval — real Basque errors from Wikipedia, hand-checked

The small, honest benchmark behind the numbers on
[`itzune/gector-eus-v5`](https://huggingface.co/itzune/gector-eus-v5),
[`itzune/gector-eus-v2`](https://huggingface.co/itzune/gector-eus-v2) and
[`itzune/latxa-8b-gec-lora-v2`](https://huggingface.co/itzune/latxa-8b-gec-lora-v2):
**171 real errors** and **247 clean sentences** taken from actual Basque Wikipedia
revisions and manually verified, so the "clean" half really is clean and the
"error" half really contains the mistake the editor fixed.

Synthetic error-injection benchmarks flatter GEC models; this one does not.

## Composition

| config | split | rows | what it is |
|---|---|---:|---|
| `default` | `errorful` | 171 | real corrections, labelled by error type |
| `default` | `clean` | 247 | sentences no editor touched (false-positive control) |
| `review` | `judged` | 600 | the mined revision pairs these came from, with human verdicts |

```python
from datasets import load_dataset
err = load_dataset("itzune/eusgec-wiki-eval", split="errorful")
clean = load_dataset("itzune/eusgec-wiki-eval", split="clean")
print(err[0])   # source, target, error_type, category, ...
```

### Error types in the `errorful` split

| type | n |
|---|---:|
| `spelling` | 52 |
| `case` | 38 |
| `punctuation` | 28 |
| `inflection` | 17 |
| `from_clean` (regression introduced by an editor) | 12 |
| `multi` | 10 |
| `word_delete` | 6 |
| `word_insert` | 5 |
| `rewrite` | 2 |
| `word_swap` | 1 |

## Fields

`errorful`: `id` (`wk-*`), `nature` (`error`), `category` (`real`), `error_type`,
`domain` (`wikipedia`), `source`, `target`.

`clean`: `sentence`.

`review/judged`: `id`, `shape` (edit shape: `spelling`, `inflection`, `word_swap`,
`rewrite`, `multi`, …), `source`, `target`, `change`, `comment` (Wikipedia edit
summary), `verdict` (`Y`/`N` — is this a real error?), `fixed_target`, `notes`
(human notes from the verification pass).

## How it was built

1. Mine sentence-level revision pairs from the Basque Wikipedia dump
   `euwiki-20261101`, keeping small edits whose corrected sentence survived to the
   latest revision.
2. Sample pairs across edit shapes; for each, render the original and corrected
   sentence and record the edit summary.
3. **Manually verify every row**: does the diff reflect a genuine error (not a
   rewrite, not vandalism, not a stylistic change)? Rows marked `N` were dropped
   from `errorful`; the `review` config keeps them so the noise rate of mined
   Wikipedia edits stays measurable.
4. `category = real`, `domain = wikipedia` distinguishes this slice from the
   synthetic horkonpon slices when pooled.

Every article that contributed here is **excluded** from
[`itzune/wiki-gec-distill-eu`](https://huggingface.co/datasets/itzune/wiki-gec-distill-eu)
and from the horkonpon Wikipedia configs, so the benchmark stays out of training.

## Metrics worth reporting

Score with `eusgec eval` (the scorer used for the published numbers) or any
M²-style scorer:

- **exact match** on `errorful` (sentence fully corrected)
- **token-level P / R / F0.5** (β = 0.5: a false correction is worse than a miss)
- **clean false-positive rate** on `clean` — the number that decides whether a
  browser assistant is usable

Published reference points (int4 ONNX GECToR unless stated):

| system | exact | P | R | F0.5 | clean FP |
|---|---:|---:|---:|---:|---:|
| GECToR v2 | — | — | — | 15.7 | 0.0% |
| GECToR v4.1 + cascade | 8.2% | 55.6 | 7.9 | 25.2 | 0.0% |
| GECToR v5 + cascade | 14.6% | 71.7 | 17.4 | 44.1 | 1.2% |
| latxa-8b-gec-lora-v2, margin −1 | 28.1% | 75.3 | 30.5 | 58.2 | 2.8% |

## Limitations

- **Small**: 171 errorful sentences. One sentence = 0.6 points of exact match;
  differences under ~2 points are noise.
- **Wikipedia register only** — encyclopedic prose, not news, not social media.
- Labels are single-annotator; the `review` verdicts are one Basque speaker's
  judgement, not a multi-annotator consensus.
- Wikipedia corrections include stylistic preferences; `from_clean` rows keep the
  handful of "corrections" that introduced an error, which most systems cannot
  catch and which depress every score slightly.

## Licensing and attribution

CC-BY-SA-4.0 — derived from the Basque Wikipedia dump `euwiki-20261101`.
The `review` config includes edit summaries and revision context from Wikipedia;
article titles are intentionally not stored, but the sentences are verbatim
Wikipedia content, so attribution to the project and share-alike both apply.

## Citation

```bibtex
@misc{ezpeleta2026eusgecwikieval,
  author       = {Xabi Ezpeleta},
  title        = {eusgec-wiki-eval: a hand-checked benchmark of real Basque
                  grammatical errors from Wikipedia},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/itzune/eusgec-wiki-eval}}
}
```

## References

- Basque Wikipedia dump `euwiki-20261101`,
  [dumps.wikimedia.org/euwiki](https://dumps.wikimedia.org/euwiki/) (CC-BY-SA-4.0)
- Scorer and baselines: [itzune/gector-eus-v2](https://github.com/itzune/gector-eus-v2)
  (`scripts/eval_onnx.py`, `eusgec eval`)
- Bryant, C., Felice, M., Briscoe, T. (2017). *Automatic Annotation and Evaluation
  of Error Types in Learner Corpora* — M² alignment convention.
- Omelianchuk, K. et al. (2020). *GECToR — Tag, Not Rewrite.* BEA@ACL.
  [arXiv:2005.12592](https://arxiv.org/abs/2005.12592)
- Evaluated systems: [`itzune/gector-eus-v5`](https://huggingface.co/itzune/gector-eus-v5),
  [`itzune/gector-eus-v2`](https://huggingface.co/itzune/gector-eus-v2),
  [`itzune/latxa-8b-gec-lora-v2`](https://huggingface.co/itzune/latxa-8b-gec-lora-v2)

	---
	license: cc-by-sa-4.0
	language:
	- eu
	pretty_name: eusgec-wiki-eval — hand-checked real Basque Wikipedia errors
	size_categories:
	- n<1K
	tags:
	- grammatical-error-correction
	- gec
	- basque
	- euskara
	- evaluation
	- wikipedia
	configs:
	- config_name: default
	data_files:
	- split: errorful
	path: default/errorful-*.parquet
	- split: clean
	path: default/clean-*.parquet
	- config_name: review
	data_files:
	- split: judged
	path: review/judged-*.parquet
	---

	# eusgec-wiki-eval — real Basque errors from Wikipedia, hand-checked

	The small, honest benchmark behind the numbers on
	[`itzune/gector-eus-v5`](https://huggingface.co/itzune/gector-eus-v5),
	[`itzune/gector-eus-v2`](https://huggingface.co/itzune/gector-eus-v2) and
	[`itzune/latxa-8b-gec-lora-v2`](https://huggingface.co/itzune/latxa-8b-gec-lora-v2):
	171 real errors and 247 clean sentences taken from actual Basque Wikipedia
	revisions and manually verified, so the "clean" half really is clean and the
	"error" half really contains the mistake the editor fixed.

	Synthetic error-injection benchmarks flatter GEC models; this one does not.

	## Composition

	\| config \| split \| rows \| what it is \|
	\|---\|---\|---:\|---\|
	\| `default` \| `errorful` \| 171 \| real corrections, labelled by error type \|
	\| `default` \| `clean` \| 247 \| sentences no editor touched (false-positive control) \|
	\| `review` \| `judged` \| 600 \| the mined revision pairs these came from, with human verdicts \|

	```python
	from datasets import load_dataset

	err = load_dataset("itzune/eusgec-wiki-eval", split="errorful")
	clean = load_dataset("itzune/eusgec-wiki-eval", split="clean")
	print(err[0]) # source, target, error_type, category, ...
	```

	### Error types in the `errorful` split

	\| type \| n \|
	\|---\|---:\|
	\| `spelling` \| 52 \|
	\| `case` \| 38 \|
	\| `punctuation` \| 28 \|
	\| `inflection` \| 17 \|
	\| `from_clean` (regression introduced by an editor) \| 12 \|
	\| `multi` \| 10 \|
	\| `word_delete` \| 6 \|
	\| `word_insert` \| 5 \|
	\| `rewrite` \| 2 \|
	\| `word_swap` \| 1 \|

	## Fields

	`errorful`: `id` (`wk-*`), `nature` (`error`), `category` (`real`), `error_type`,
	`domain` (`wikipedia`), `source`, `target`.

	`clean`: `sentence`.

	`review/judged`: `id`, `shape` (edit shape: `spelling`, `inflection`, `word_swap`,
	`rewrite`, `multi`, …), `source`, `target`, `change`, `comment` (Wikipedia edit
	summary), `verdict` (`Y`/`N` — is this a real error?), `fixed_target`, `notes`
	(human notes from the verification pass).

	## How it was built

	1. Mine sentence-level revision pairs from the Basque Wikipedia dump
	`euwiki-20261101`, keeping small edits whose corrected sentence survived to the
	latest revision.
	2. Sample pairs across edit shapes; for each, render the original and corrected
	sentence and record the edit summary.
	3. Manually verify every row: does the diff reflect a genuine error (not a
	rewrite, not vandalism, not a stylistic change)? Rows marked `N` were dropped
	from `errorful`; the `review` config keeps them so the noise rate of mined
	Wikipedia edits stays measurable.
	4. `category = real`, `domain = wikipedia` distinguishes this slice from the
	synthetic horkonpon slices when pooled.

	Every article that contributed here is excluded from
	[`itzune/wiki-gec-distill-eu`](https://huggingface.co/datasets/itzune/wiki-gec-distill-eu)
	and from the horkonpon Wikipedia configs, so the benchmark stays out of training.

	## Metrics worth reporting

	Score with `eusgec eval` (the scorer used for the published numbers) or any
	M²-style scorer:

	- exact match on `errorful` (sentence fully corrected)
	- token-level P / R / F0.5 (β = 0.5: a false correction is worse than a miss)
	- clean false-positive rate on `clean` — the number that decides whether a
	browser assistant is usable

	Published reference points (int4 ONNX GECToR unless stated):

	\| system \| exact \| P \| R \| F0.5 \| clean FP \|
	\|---\|---:\|---:\|---:\|---:\|---:\|
	\| GECToR v2 \| — \| — \| — \| 15.7 \| 0.0% \|
	\| GECToR v4.1 + cascade \| 8.2% \| 55.6 \| 7.9 \| 25.2 \| 0.0% \|
	\| GECToR v5 + cascade \| 14.6% \| 71.7 \| 17.4 \| 44.1 \| 1.2% \|
	\| latxa-8b-gec-lora-v2, margin −1 \| 28.1% \| 75.3 \| 30.5 \| 58.2 \| 2.8% \|

	## Limitations

	- Small: 171 errorful sentences. One sentence = 0.6 points of exact match;
	differences under ~2 points are noise.
	- Wikipedia register only — encyclopedic prose, not news, not social media.
	- Labels are single-annotator; the `review` verdicts are one Basque speaker's
	judgement, not a multi-annotator consensus.
	- Wikipedia corrections include stylistic preferences; `from_clean` rows keep the
	handful of "corrections" that introduced an error, which most systems cannot
	catch and which depress every score slightly.

	## Licensing and attribution

	CC-BY-SA-4.0 — derived from the Basque Wikipedia dump `euwiki-20261101`.
	The `review` config includes edit summaries and revision context from Wikipedia;
	article titles are intentionally not stored, but the sentences are verbatim
	Wikipedia content, so attribution to the project and share-alike both apply.

	## Citation

	```bibtex
	@misc{ezpeleta2026eusgecwikieval,
	author = {Xabi Ezpeleta},
	title = {eusgec-wiki-eval: a hand-checked benchmark of real Basque
	grammatical errors from Wikipedia},
	year = {2026},
	publisher = {Hugging Face},
	howpublished = {\url{https://huggingface.co/datasets/itzune/eusgec-wiki-eval}}
	}
	```

	## References

	- Basque Wikipedia dump `euwiki-20261101`,
	[dumps.wikimedia.org/euwiki](https://dumps.wikimedia.org/euwiki/) (CC-BY-SA-4.0)
	- Scorer and baselines: [itzune/gector-eus-v2](https://github.com/itzune/gector-eus-v2)
	(`scripts/eval_onnx.py`, `eusgec eval`)
	- Bryant, C., Felice, M., Briscoe, T. (2017). *Automatic Annotation and Evaluation
	of Error Types in Learner Corpora* — M² alignment convention.
	- Omelianchuk, K. et al. (2020). GECToR — Tag, Not Rewrite. BEA@ACL.
	[arXiv:2005.12592](https://arxiv.org/abs/2005.12592)
	- Evaluated systems: [`itzune/gector-eus-v5`](https://huggingface.co/itzune/gector-eus-v5),
	[`itzune/gector-eus-v2`](https://huggingface.co/itzune/gector-eus-v2),
	[`itzune/latxa-8b-gec-lora-v2`](https://huggingface.co/itzune/latxa-8b-gec-lora-v2)