OmniGEC: Multilingual GEC Dataset

• OmniGEC is a multilingual silver-standard dataset for paragraph-level grammatical error correction built from Wikipedia edits, Reddit posts, and Ukrainian social media texts.
• It addresses non-English data scarcity by standardizing corrections in 11 languages with both human editing and synthetic pipelines using tools like GPT-4o-mini.
• Experimental results demonstrate state-of-the-art performance in low-resource languages and improved fluency with tailored multilingual LLM fine-tuning.

Searching arXiv for the cited OmniGEC paper and closely related multilingual GEC work to ground the article. OmniGEC is a multilingual silver-standard dataset collection for grammatical error correction (GEC) introduced to address the shortage of parallel correction data outside English and to support paragraph-level multilingual GEC with modern LLMs. It covers eleven languages—Czech, English, Estonian, German, Greek, Icelandic, Italian, Latvian, Slovene, Swedish, and Ukrainian—and combines human-derived Wikipedia edits with synthetically corrected Reddit and Ukrainian social-media text. The resource is explicitly positioned as infrastructure for moving beyond English-centric GEC, and the associated experiments report state-of-the-art results for paragraph-level multilingual GEC under specific training settings (Kovalchuk et al., 18 Sep 2025).

1. Definition, scope, and research setting

OmniGEC is a collection of multilingual silver-standard corpora designed for GEC at the paragraph level. Its central motivation is that multilingual GEC lags behind English because parallel data is expensive, gold data is scarce, datasets are not standardized across languages, and many methods remain English-specific. The paper presents OmniGEC as a response to the mismatch between the needs of modern LLM-based GEC and the limited scale of existing multilingual corpora (Kovalchuk et al., 18 Sep 2025).

The collection is aligned with the eleven-language setup of MultiGEC-2025 so that data and models can be compared directly. The paper emphasizes that existing multilingual resources are valuable but too small and heterogeneous for large-scale fine-tuning, ablation studies, or robust multilingual comparison. OmniGEC is therefore intended to provide higher-volume multilingual training data, support ablations across silver sources, reduce dependence on English-centric assumptions, enable adaptation of open-source multilingual LLMs to paragraph-level GEC, and create a foundation that can be extended with new languages and new samples over time (Kovalchuk et al., 18 Sep 2025).

A key terminological point is that all three components are treated as silver-standard. This includes the Wikipedia-derived portion, even though it comes from human edits, because those edits may contain non-GEC changes such as information updates, formatting, or context distortion. The paper therefore uses “silver-standard” as a task-specific quality designation rather than as a synonym for automatically generated data (Kovalchuk et al., 18 Sep 2025).

2. Corpus composition and data organization

OmniGEC consists of three subcorpora with distinct provenance and correction mechanisms. WikiEdits-MultiGEC is based on human-made Wikipedia edits in all eleven languages. Reddit-MultiGEC contains multilingual Reddit paragraphs whose corrections were generated automatically. UberText-GEC is a Ukrainian-only synthetic corpus derived from UberText 2.0 social-media text (Kovalchuk et al., 18 Sep 2025).

Subcorpus	Coverage	Source/correction type
WikiEdits-MultiGEC	11 languages	Wikipedia edits; human-derived corrections
Reddit-MultiGEC	11 languages	Reddit paragraphs; automatically corrected
UberText-GEC	Ukrainian only	UberText 2.0 paragraphs; automatically corrected

The dataset is organized by language and split into train, development, and test sets using an 80/10/10 ratio per language. The texts are handled at the paragraph level rather than the sentence level, which directly shapes both the construction pipeline and the downstream modeling experiments (Kovalchuk et al., 18 Sep 2025).

The clearest explicit corpus-scale statistics reported in the paper are the following. WikiEdits-MultiGEC averages about 1.6K samples per language and totals 1.2M tokens. Reddit-MultiGEC totals 13M tokens. UberText-GEC uses a 25% subset of UberText 2.0, corresponding to 22M tokens out of 110M total (Kovalchuk et al., 18 Sep 2025).

The paper also reports readable sample counts for selected language-source combinations. In WikiEdits, visible counts include English 6,807, Ukrainian 1,872, German 2,386, Greek 3,142, and Estonian 2,568. In Reddit, visible counts include English 51,671, German 44,551, Czech 17,891, Estonian 10,673, Greek 7,984, Slovene 9,122, Latvian 6,836, and Ukrainian 5,032. UberText-GEC contains 199,950 Ukrainian samples in the appendix figure. The paper notes that some appendix values are partially garbled in extraction, so not every language/source count is recoverable from the provided text (Kovalchuk et al., 18 Sep 2025).

3. Construction methodology and synthetic correction pipeline

The three subcorpora differ substantially in how source-target pairs are created. In WikiEdits-MultiGEC, the source is pre-edit Wikipedia text and the target is post-edit text. The source data was collected through the official Wikipedia API over a six-month period from September 28, 2024 to April 17, 2025, and restricted to the “newcomer task copyedit” category because those edits tend to contain small grammatical corrections rather than broad content revisions (Kovalchuk et al., 18 Sep 2025).

The Wikipedia construction process had three main steps: collection of metadata for recently edited pages, collection of all edits from each page, and post-processing to remove Wikipedia-specific artifacts. Filtering rules excluded samples shorter than 50 characters, samples with more than 10 corrections, and samples beginning with special characters such as ==, !, and |. The pipeline also cleaned referral links, citations, code tags, language-specific tags, and related formatting artifacts. English was artificially capped to avoid intensifying English-resource dominance (Kovalchuk et al., 18 Sep 2025).

Reddit-MultiGEC and UberText-GEC use a synthetic correction pipeline built around DeepL, o1-preview, and GPT-4o-mini. The Reddit component was collected from subreddits whose main language matched one of the eleven target languages, with an additional Icelandic-learning subreddit included for Icelandic because of its extremely low-resource status. Before correction, Reddit texts passed through two filters: language identification with langid and content moderation with OpenAI’s omni-moderation-2024-09-26. The highest censored fraction was Italian at nearly 20%, while the lowest was Icelandic at 2.8%. The Reddit collection was capped at the latest 400 subreddits per language as of March 25, 2025 (Kovalchuk et al., 18 Sep 2025).

The synthetic pipeline is described as a three-step process. First, the authors wrote a GEC instruction in English, translated it into each target language using DeepL, extracted correction examples from the development set of MultiGEC, and asked o1-preview to generate a language-specific few-shot prompt. Second, GPT-4o-mini was used to generate three paragraph-level corrections per sample. Third, GPT-4o-mini was prompted again to aggregate the three candidate corrections into one final correction. The aggregation prompt was also originally written in English and translated into the target languages using DeepL (Kovalchuk et al., 18 Sep 2025).

The paper frames this aggregation design as a variation on high-diversity ranking or ensembling from prior English GEC work, but with an important change: instead of selecting one candidate, OmniGEC merges multiple candidates into a single final output. The stated reason is that GPT-4o-mini often distributes edits across alternatives, so aggregation can produce a more complete correction. This also introduces a known risk of overcorrection, which the authors discuss explicitly (Kovalchuk et al., 18 Sep 2025).

4. Quality assessment and corpus-specific error profiles

Quality evaluation is reported in both automatic and human forms, but only for Ukrainian. The paper identifies this language restriction as a limitation, since the other ten languages do not receive direct human validation in the reported study (Kovalchuk et al., 18 Sep 2025).

For automatic evaluation, the authors created proxy references using three publicly available Ukrainian GEC systems: Pravopysnyk, Spivavtor, and LanguageTool. They sampled 1,500 Ukrainian examples from each of the three subcorpora, yielding 4,500 total samples, and compared OmniGEC corrections against the three reference outputs using ERRANT precision, recall, and $F_{0.5}$, GLEU, Levenshtein distance, and character error rate (Kovalchuk et al., 18 Sep 2025).

Corpus	Precision	Recall	$F_{0.5}$
Reddit-MultiGEC	17.92	59.51	20.84
UberText-GEC	16.83	56.81	19.59
WikiEdits-MultiGEC	13.30	26.03	14.74

The same table also reports GLEU, Levenshtein distance, and character error rate: Reddit-MultiGEC has 46.89 GLEU, 36.87 Levenshtein distance, and 18.20 character error rate; UberText-GEC has 63.45 GLEU, 23.51 Levenshtein distance, and 10.98 character error rate; WikiEdits-MultiGEC has 71.35 GLEU, 18.21 Levenshtein distance, and 4.79 character error rate (Kovalchuk et al., 18 Sep 2025). The authors interpret these results as showing that as character error rate increases, GLEU decreases and $F_{0.5}$ increases. In their reading, corpora with more edits diverge more from proxy references in overlap-based metrics such as GLEU while producing higher edit-oriented ERRANT scores.

Human evaluation was conducted as a grading task on the same 4,500 Ukrainian samples. Annotators scored each correction from 1 to 5, where 4 corresponds to a correction that conforms to Ukrainian orthography and is labeled the “minimal” grade, and 5 corresponds to a correction that conforms to orthography and improves fluency, labeled the “fluency” grade. A reject option was available for incomprehensible originals or impossible-to-judge corrections. The annotation team comprised 15 native Ukrainian speakers, mostly linguistics students. Only 100 samples were double-annotated, and no inter-annotator agreement coefficient is reported (Kovalchuk et al., 18 Sep 2025).

Average human grades were 3.05 for WikiEdits-MultiGEC, 3.52 for Reddit-MultiGEC, and 3.66 for UberText-GEC. Reject rates were 9.9% for WikiEdits-MultiGEC, 2.8% for Reddit-MultiGEC, and 2.3% for UberText-GEC. The paper further states that more than 70% of examples in the synthetic corpora received grades 4 or 5. One of the paper’s most notable empirical observations is therefore that the synthetic GPT-4o-mini corrections were rated better on average than the harvested Wikipedia edits (Kovalchuk et al., 18 Sep 2025).

The error analysis clarifies why these averages differ. Across corpora, low grades were associated with residual errors, overcorrection, and texts with too many corrections to judge comfortably. In Reddit-MultiGEC, low grades sometimes reflected non-ethical or inappropriate content that persisted despite moderation. In UberText-GEC, low grades often involved insertion of extra non-essential text such as promotional phrases. In WikiEdits-MultiGEC, failure modes included information updates rather than grammar correction, domain-specific corrections requiring specialist knowledge, distortion of context through deletions or additions, and leftover formatting artifacts. This analysis directly challenges the assumption that human-derived edits are automatically better suited to GEC than synthetic ones (Kovalchuk et al., 18 Sep 2025).

5. Modeling experiments and empirical performance

The paper fine-tunes two open-source multilingual LLMs on combinations of MultiGEC and OmniGEC data: Aya-Expanse 8B and Gemma-3-12B-IT. Experiments follow the MultiGEC-2025 shared-task setting and evaluate both the minimal track and the fluency track. Even though the synthetic prompts targeted the minimal setting, the models were also evaluated on fluency because human judgments indicated that 7–9% of examples improved fluency as well (Kovalchuk et al., 18 Sep 2025).

Three training conditions were used for each model and each track: training only on MultiGEC; training on MultiGEC plus WikiEdits-MultiGEC; and training on MultiGEC plus WikiEdits-MultiGEC plus Reddit-MultiGEC. UberText-GEC was not included in training because of time and cost limitations. The fine-tuning setup was multilingual and paragraph-level throughout (Kovalchuk et al., 18 Sep 2025).

The main comparison point is Lattice, the best paragraph-level solution submitted to MultiGEC-2025, which fine-tuned LLaMA-3 8B. Evaluation was conducted on the MultiGEC-2025 test set in the CodaLab environment, and the primary reported metric is GLEU (Kovalchuk et al., 18 Sep 2025).

Mean GLEU over all languages shows distinct behaviors for the two models. Aya-Expanse-8B scores 64.52 minimal and 48.37 fluency when trained only on MultiGEC, 65.16 and 48.37 with WikiEdits added, and 65.43 and 49.80 with WikiEdits plus Reddit. Gemma-3-12B-IT scores 61.43 and 48.66 on MultiGEC alone, 67.02 and 52.34 with WikiEdits added, and 66.42 and 49.20 with WikiEdits plus Reddit (Kovalchuk et al., 18 Sep 2025). The paper’s narrative conclusion is that both Aya-Expanse and Gemma-3, when fine-tuned with OmniGEC data, outperform the previous leading paragraph-based multilingual editing model.

The paper states several explicit highlight results. For Ukrainian fluency in the paragraph-based setting, Aya-Expanse-8B trained on MultiGEC+Wiki reaches 77.10 GLEU and is described as the best score for paragraph-based edits. For Latvian paragraph-based editing, Gemma-3-12B-IT trained on MultiGEC+Wiki reaches 81.54 GLEU, which the paper also describes as the best score for paragraph-based edits. For Estonian minimal edits, Gemma-3-12B-IT trained on MultiGEC+Wiki+Reddit reaches 57.54 GLEU, and the paper states that this constitutes a state-of-the-art result on the Estonian minimal edits track (Kovalchuk et al., 18 Sep 2025).

The ablation results are informative rather than uniformly monotonic. More silver data generally improves mean performance, but gains vary by language and source. WikiEdits alone sometimes outperforms WikiEdits plus Reddit, especially for Gemma-3 under the reported training budget. The authors suggest two reasons: data-quality differences and undertraining, since the Reddit-augmented condition is much larger and Gemma-3 was trained for only one epoch (Kovalchuk et al., 18 Sep 2025). A plausible implication is that the paper’s ablations combine effects of corpus composition and optimization budget rather than isolating corpus quality alone.

The paper interprets the model behavior in resource-sensitive terms. Aya-Expanse 8B is stronger when trained only on the small gold MultiGEC baseline, whereas Gemma-3 12B benefits much more from added silver data. Lower-resource languages appear to gain most from OmniGEC augmentation. The authors specifically highlight a 26.14 GLEU gain for Gemma on Latvian over its own baseline, and for Aya the largest gain on Estonian, at +8.25 minimal and +4.97 fluency (Kovalchuk et al., 18 Sep 2025).

6. Training configuration, limitations, and disambiguation

The appendix reports training configurations for both models. For Aya-Expanse-8B, inference used temperature 0.3 and top_p 0.75; training used learning rate $3\mathrm{e}{-5}$, optimizer paged_adamw_32bit, gradient checkpointing true, bf16 true, cosine learning-rate scheduling, early stopping, and LoRA target modules q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj. For Gemma-3-12B-IT, inference used temperature 1.0 and top_p 0.95; training used learning rate $3\mathrm{e}{-5}$, optimizer adamw_torch_fused, gradient checkpointing true, bf16 true, cosine scheduling, LoRA target modules all-linear, and modules_to_save: lm_head, embed_tokens. Shared settings include task type CAUSAL_LM, weight decay 0.0, max grad norm 1.0, save/eval by steps, best-model selection by eval_loss, and loading the best model at the end (Kovalchuk et al., 18 Sep 2025).

The hardware note is also consequential for interpreting results. Gemma-3-12B-IT was trained for only one epoch because it took almost a day on a single A100 40GB GPU even with packing and batching, whereas Aya-Expanse-8B completed three epochs in the same 24-hour window before reaching a plateau. This reported asymmetry is one reason the paper treats some of its ablation findings cautiously (Kovalchuk et al., 18 Sep 2025).

The paper is explicit about several limitations. OmniGEC covers only eleven languages. Human evaluation is only reported for Ukrainian. The synthetic generation pipeline depends on proprietary models, specifically GPT-4o-mini and o1-preview, which affects reproducibility. Only two open-source LLMs were explored. The silver data may contain overcorrections, especially because the pipeline aggregates multiple generated candidates into one. Wikipedia edits are noisy for GEC because they include factual updates, formatting artifacts, and domain-specific rewrites. Synthetic corrections may insert extra content, alter style unnecessarily, or over-edit. Reddit data may still contain harmful content, misinformation, or bias despite moderation, and moderation quality was not qualitatively assessed across all languages (Kovalchuk et al., 18 Sep 2025).

The release information is concrete but incomplete in one respect. The dataset collection and the best-performing models are available on Hugging Face, and data collection scripts, prompting templates, and model code are released on GitHub. The paper excerpt does not state a formal software or data license (Kovalchuk et al., 18 Sep 2025).

A final point of clarification concerns nomenclature. “OmniGEC” in this context refers to the multilingual grammatical error correction dataset and associated experiments introduced in 2025 (Kovalchuk et al., 18 Sep 2025). It should not be conflated with “OmniJet-$\alpha_C$,” which is a calorimeter simulation model in high-energy physics (Birk et al., 9 Jan 2025), or with “Omni” / “Omni-GeoEncoder,” which refers to a geospatial entity resolution architecture (Wijegunarathna et al., 8 Aug 2025). The shared “Omni” prefix spans unrelated research programs; the GEC resource is distinct in task, data modality, and methodology.

The future directions outlined for OmniGEC include evaluating more models, exploring sentence-based editing, using preference optimization such as DPO with the human score annotations, conducting more thorough quantity-versus-quality ablations, especially with UberText-GEC, testing per-language LoRA adapters, improving Ukrainian GEC further, and continuously expanding the dataset with additional data and languages (Kovalchuk et al., 18 Sep 2025). In that sense, OmniGEC is best understood not as a fixed benchmark but as a multilingual silver-data infrastructure for paragraph-level GEC research.