GPT model : Benchmark

Content

The following repository consists of a benchmark of Language Models (GPT-3.5 : text-davinci-003, BERT, and GPT-2) in the case of text summarization task of 115 sentences from the test dataset of the CNN-DailyMail News from Kaggle (https://www.kaggle.com/datasets/gowrishankarp/newspaper-text-summarization-cnn-dailymail).

To use ChatGPT

Before running the scripts, and to be able to use ChatGPT for the experiments, you have to put your api key on openai_key.txt to be able to use the OpenAI models. To get it, you must an OpenAI account available by doing the following steps:

1. Navigate to https://platform.openai.com/
2. Click on your avatar in the top right-hand corner of the dashboard.
3. Select View API Keys.
4. Click Create new secret key.

API Key Good Practice Safety are available on OpenAI website : https://help.openai.com/en/articles/5112595-best-practices-for-api-key-safety

For the moment, only BERT, GPT-2 and text-davinci-003 (fine-tuned GPT-3.5) models are evaluated.

How to run

Located at the root of the project, you can generate the summarized sentences by writing in your terminal:

python experiments/summarization_task.py

On the following repository, the results are given as results/generated_summary.txt. To get the average value of BLEU, ROUGE and BLEURT metrics over the 115 sentences summarized, you have to write:

python experiments/benchmark.py

Creating a table : model_summarization_score.csv in results folder.

Requirements

Python 3.11.2

• bert-extractive-summarizer==0.10.1
• openai==0.27.4
• rouge==1.0.1
• tensorflow==2.12.0
• tensorflow-datasets==4.9.2
• torch==2.0.0
• transformers==4.28.1

You also have to install BLEURT metric from https://github.com/google-research/bleurt#readme such as they stated:

pip install --upgrade pip  # ensures that pip is current
git clone https://github.com/google-research/bleurt.git
cd bleurt
pip install .