arabert-base-v02-finetuned-MRC

arabert-base-v02-finetuned-MRC

Model Description

This model is a fine-tuned version of bert-base-arabertv02 on an expert-annotated Arabic Machine Reading Comprehension (MRC) dataset <a href="https://huggingface.co/datasets/RashaMObeidat/ArQuAD">ArQuAD</a> , for the MRC task.

• Model type: Question answering model
• Downstream-task: Extractive QA
• Language(s) (NLP): Arabic
• Finetuned from model : <a href="https://huggingface.co/aubmindlab/bert-base-arabertv02">aubmindlab/bert-base-arabertv02</a>

How to Get Started with the Model

Use the code below to get started with the model:

model = AutoModelForQuestionAnswering.from_pretrained("AyaZJaradat/arabert-base-v02-finetuned-MRC")
tokenizer = AutoTokenizer.from_pretrained("AyaZJaradat/arabert-base-v02-finetuned-MRC")

Training Details

Training Data

We use the ArQuAD dataset to fine-tune the model, with 12,816 rows of training.

Preprocessing

We removed the diacritics in the data before starting training.

Training Hyperparameters

• Learning Rate: 3e-5
• Optimizer: AdamW
• Batch size: 12
• Epochs: 3

Evaluation

Testing Data

We used the testing data from ArQuAD with 1602 rows.

Metrics

We used the Exact Match (EM) and the F1-score to evaluate the model using the script of SQuAD.

Results

The results for the test data are as follows:

• EM: 66.73%
• F1-score: 85.52%

Model Card Authors

Aya Jaradat

Model Card Contact

[email protected]