RomanCast/WikiSpell_custom

WikiSpell

Description

This dataset is a custom implementation of the WikiSpell dataset introduced in Character-Aware Models Improve Visual Text Rendering by Liu et al. (2022).

Similarly to the original WikiSpell dataset, the training set is composed of 5000 words taken uniformly from the 50% least common Wiktionary words (taken from this Wiktionary extraction), and 5000 words sampled according to their frequencies taken from the 50% most common Wiktionary words.

The validation and test are splitted in 5 sets, sampled depending on their frequency in the corpus:

• 1% most common words
• 1 - 10% most common words
• 10 - 20% most common words
• 20 - 30% most common words
• 50% least common words

Contrary to the original WikiSpell dataset, we compute the frequency of the words using the first 100k sentences from OpenWebText (Skylion007/openwebtext) instead of mC4.

Usage

This dataset is used for testing spelling in Large Language Models. To do so, the labels should be computed like in the following snippet:

sample = ds["train"][0]
label = " ".join(sample["text"])

The labels are not included in the dataset files directly.

Citation

Please cite the original paper introducing WikiSpell if you're using this dataset:

@inproceedings{liu-etal-2023-character,
    title = "Character-Aware Models Improve Visual Text Rendering",
    author = "Liu, Rosanne  and
      Garrette, Dan  and
      Saharia, Chitwan  and
      Chan, William  and
      Roberts, Adam  and
      Narang, Sharan  and
      Blok, Irina  and
      Mical, Rj  and
      Norouzi, Mohammad  and
      Constant, Noah",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.900",
    pages = "16270--16297",
}