CoolFace
Modelpublic

salti/arabic-t5-small-question-paraphrasing

sourceHugging Faceupdated 3y agoView on Hugging Face
3likes59downloads
Model Card

Arabic T5v1.1 for question paraphrasing

This is a fine-tuned arabic-t5-small on the task of question paraphrasing.

A demo of the trained model using HF Spaces can be found here

Training data

The model was fine-tuned using the Semantic Question Similarity in Arabic data on kaggle.

Only the rows of the dataset where the label is True (the two questions have the same meaning) were taken.

The training data was then also mirrored; so if q1 and q2 were two questions with the same meaning, then (q1, q2) and (q2, q1) were both present in the training set. The evaluation set was kept unmirrored of course.

Training config

batch size128
dropout rate0.1
learning rate0.001
lr scheduleconstant
weight decay1e-7
epochs3

Results

training loss0.7086
evaluation loss0.9819
meteor49.277
sacreBLEU-157.088
sacreBLEU-239.846
sacreBLEU-329.444
sacreBLEU-422.601
Rouge F1 max1.299