metadata
language: uk
tags:
- gec
- mbart-50
widget:
- text: я й не думав що комп'ютерна лінгвістика це легкоо.
This model was finetuned on errorful sentences from the train
subset of UA-GEC corpus, introduced in UA-GEC: Grammatical Error Correction and Fluency Corpus for the Ukrainian Language paper.
Only sentences containing errors were used; 8,874 sentences for training and 987 sentences for validation. The training arguments were defined as follows:
batch_size = 4
num_train_epochs = 3
learning_rate=5e-5
weight_decay=0.01
optim = "adamw_hf"