About the demo

In this demo you can work with the models for grammatical error correction in English. There are two models available - T5-Efficient-MINI (31M parameters) and T5-Efficient-TINY (16M parameters). Both models were trained on a portion (due to the time constraints) of the C4_200M dataset (37M training samples and 37M validation samples). In addition to the original errors of the dataset, more basic typos were dynamically introduced to the training set with the help of nlpaug library. As a result, T5-Efficient-TINY reached validation loss of 0.08 and T5-Efficient-MINI achieved validation loss of 0.06.

Both models are available in original and quantinized variants. Models export to ONNX format was performed using fastT5 library. Exported models were quantinized using standard PyTorch functionality.

How to use the demo:

Load the model and start typing. The model will run as soon as you stop typing.