Taqyim تقييم

A library for evaluting Arabic NLP datasets on chatgpt models.

Installation

pip install -e .

Example

import taqyim as tq
pipeline = tq.Pipeline(
    eval_name="ajgt-test",
    dataset_name="arbml/ajgt_ubc_split",
    task_class="classification",
    task_description= "Sentiment Analysis",
    input_column_name="content",
    target_column_name="label",
    prompt="Predict the sentiment",
    api_key="<openai-key>",
    train_split="train",
    test_split="test",
    model_name="gpt-3.5-turbo-0301",
    max_samples=1,)

# run the evaluation
pipeline.run()

# show the output data frame
pipeline.show_results()

# show the eval metrics
pipeline.get_final_report()

Run on custom dataset

custom_dataset.ipynb has a complete example on how to run evaluation on a custom dataset.

parameters

• eval_name choose an eval name
• task_class class name from supported class names
• task_description short description about the task
• dataset_name dataset name for evaluation
• subset If the dataset has subset
• train_split train split name in the dataset
• test_splittest split name in the dataset
• input_column_name input column name in the dataset
• target_column_name target column name in the dataset
• prompt the prompt to be fed to the model
• task_description short string explaining the task
• api_key api key from keys
• preprocessing_fn function used to process inputs and targets
• threads number of threads used to fetch the api
• threads_timeout thread timeout
• max_samples max samples used for evaluation from the dataset
• model_name choose either gpt-3.5-turbo-0301 or gpt-4-0314
• temperature temperature passed to the model between 0 and 2, higher temperature means more random results
• num_few_shot number of fewshot samples to be used for evaluation
• resume_from_record if True it will continue the run from the sample that has no results.
• seed seed to redproduce the results

Supported Classes and Tasks

• Classification classification tasks see classification.py.
• Pos_Tagging part of speech tagging tasks pos_tagging.py.
• Translation machine translation translation.py.
• Summarization machine translation summarization.py.
• MCQ multiple choice question answering mcq.py.
• Rating rating multiple LLMs outputs rating.py.
• Diacritization machine translation diacritization.py.

Evaluation on Arabic Tasks

✢ BOLT requires LDC subscription

✢✢ WikiNews not public, contact authors to access the dataset

@misc{alyafeai2023taqyim,
      title={Taqyim: Evaluating Arabic NLP Tasks Using ChatGPT Models}, 
      author={Zaid Alyafeai and Maged S. Alshaibani and Badr AlKhamissi and Hamzah Luqman and Ebrahim Alareqi and Ali Fadel},
      year={2023},
      eprint={2306.16322},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}