|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- gec |
|
|
- grammar |
|
|
size_categories: |
|
|
- 10K<n<100K |
|
|
configs: |
|
|
- config_name: gec |
|
|
data_files: |
|
|
- split: train |
|
|
path: train/* |
|
|
- split: dev |
|
|
path: dev/* |
|
|
- split: dev1 |
|
|
path: dev1/* |
|
|
--- |
|
|
|
|
|
# Grammar Correctable Texts |
|
|
|
|
|
These are collected texts sampled from various sources, which have been screened across several dimensions using google flash 2.0: |
|
|
|
|
|
- `is_english` |
|
|
- `meaning_is_recoverable` (Intended meaning is clear enough that correction won't change it.) |
|
|
- `coherent_and_on_topic` (Sentences cohere; not random fragments or spam.) |
|
|
- `not_style_or_dialect_intent` (Oddities are not intentional dialect/poetry/stylized voice.) |
|
|
- `error_density_not_extreme` |
|
|
- `not_code_or_markup_heavy` |
|
|
- `complete_sentences_present` |
|
|
|
|
|
Additionally, I use it to estimate the number of errors, bin into a 2d histogram with `num_tokens` and `estimated_corrections`. |
|
|
I downsampled over-represented bins to help create a more uniform distribution of lengths and error rates. |
|
|
|
|
|
This dataset has no ground truth, and targets GRPO training for GEC (grammatical error correction). |
|
|
|
