martinsr/wi_locness

W&I+LOCNESS — BEA-2019 GEC Dataset

A reformatted version of the Write & Improve + LOCNESS corpus released as part of the BEA-2019 Shared Task on Grammatical Error Correction. This dataset is intended for non-commercial research use only, in accordance with the original license from Cambridge.

Dataset Description

The dataset combines two corpora:

Write & Improve (W&I) is an online platform where non-native English students submit writing in response to prompts and receive instant feedback. A subset of these submissions was manually annotated by Cambridge English annotators and assigned CEFR proficiency levels (A, B, C).

LOCNESS consists of essays written by native English students, originally compiled at the Centre for English Corpus Linguistics at the University of Louvain. A subset was annotated by the same W&I annotators to enable evaluation across the full range of English proficiency levels.

Format

Each row contains:

Column	Type	Description
text	string	Original (possibly erroneous) source text
edits	dict	Character-level edits with keys start, end, text (each a list)
cefr	string	CEFR proficiency level (A, B, C, or N for native)
id	string	Original submission ID

The edits dict uses the first annotator's annotations. None values in the text list indicate deletions (the span should be removed).

Applying edits

def apply_edits(text: str, edits: dict) -> str:
    if not edits or not edits.get("start"):
        return text
    edit_list = sorted(
        zip(edits["start"], edits["end"], edits["text"]),
        key=lambda x: x[0], reverse=True
    )
    corrected = text
    for start, end, rep in edit_list:
        rep = rep if rep is not None else ""  # None = deletion
        corrected = corrected[:start] + rep + corrected[end:]
    return corrected

Splits

Split	Texts	Notes
train	3,000	A/B/C levels only — no native training set was released
validation	350	A/B/C levels + 50 native (N) essays from LOCNESS

After applying edits and filtering identical source/target pairs, the training split yields approximately 2,997 correction pairs.

Provenance

Original data: W&I+LOCNESS v2.1, released by the University of Cambridge for the BEA-2019 Shared Task. Edits were standardised with ERRANT and originally distributed in M2 and JSON formats.

This version converts the original JSON format to parquet with character-level edit offsets preserved, making it directly usable with HuggingFace datasets without requiring ERRANT or M2 tooling.

Usage

from datasets import load_dataset

ds = load_dataset("martinsr/wi_locness", "wi", split="train")

# Apply edits to get correction pairs
def apply_edits(text, edits):
    if not edits or not edits.get("start"):
        return text
    edit_list = sorted(
        zip(edits["start"], edits["end"], edits["text"]),
        key=lambda x: x[0], reverse=True
    )
    corrected = text
    for start, end, rep in edit_list:
        rep = rep if rep is not None else ""
        corrected = corrected[:start] + rep + corrected[end:]
    return corrected

for row in ds:
    src = row["text"]
    tgt = apply_edits(src, row["edits"])
    if src != tgt:
        print(f"SRC: {src[:80]}")
        print(f"TGT: {tgt[:80]}")
        break

License

This dataset may only be used for non-commercial research purposes, in accordance with the original license from Cambridge English / University of Cambridge. See the BEA-2019 shared task page for full license terms.

Citation

If you use this dataset, please cite the original BEA-2019 shared task paper:

@inproceedings{bryant-etal-2019-bea,
    title     = "The {BEA}-2019 Shared Task on Grammatical Error Correction",
    author    = "Bryant, Christopher and Felice, Mariano and
                 Andersen, Øistein E. and Briscoe, Ted",
    booktitle = "Proceedings of the Fourteenth Workshop on Innovative Use
                 of NLP for Building Educational Applications",
    year      = "2019",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/W19-4406",
}

Related

• BEA-2019 Shared Task
• ERRANT — error annotation toolkit used to standardise the corpus
• martinsr/gemma-3-1b-gec-esl — GEC model trained on this dataset
• martinsr/lfm2.5-1.2b-gec-esl — GEC model trained on this dataset