nmatheg

nmatheg نماذج an easy straregy for training Arabic NLP models on huggingface datasets. Just specifiy the name of the dataset, preprocessing, tokenization and the training procedure in the config file to train an nlp model for that task.

install

pip install nmatheg

Configuration

Setup a config file for the training strategy.

[dataset]
dataset_name = ajgt_twitter_ar

[preprocessing]
segment = False
remove_special_chars = False
remove_english = False
normalize = False
remove_diacritics = False
excluded_chars = []
remove_tatweel = False
remove_html_elements = False
remove_links = False 
remove_twitter_meta = False
remove_long_words = False
remove_repeated_chars = False

[tokenization]
tokenizer_name = WordTokenizer
vocab_size = 1000
max_tokens = 128

[model]
model_name = rnn

[log]
print_every = 10

[train]
save_dir = .
epochs = 10
batch_size = 256 

Main Sections

• dataset describe the dataset and the task type. Currently we only support classification
• preprocessing a set of cleaning functions mainly uses our library tnkeeh.
• tokenization descrbies the tokenizer used for encoding the dataset. It uses our library tkseem.
• train the training parameters like number of epochs and batch size.

Usage

Config Files

import nmatheg as nm
strategy = nm.TrainStrategy('config.ini')
strategy.start()

Benchmarking on multiple datasets and models

import nmatheg as nm
strategy = nm.TrainStrategy(
    datasets = 'arsentd_lev,caner,arcd', 
    models   = 'qarib/bert-base-qarib,aubmindlab/bert-base-arabertv01',
    mode = 'finetune',
    runs = 5,
    lr = 1e-4,
    epochs = 1,
    batch_size = 8,
    max_tokens = 128,
    max_train_samples = 1024
)
strategy.start()

Datasets

We are supporting huggingface datasets for Arabic. You can find the supported datasets here.

Tasks

Currently we support text classification, named entity recognition, question answering, machine translation and natural language inference.

Demo

Check this colab notebook for a quick demo.