GreekT5 umt5 base greeksum EU AI Act Compliance Profile

GreekT5 (umt5-base-greeksum)

A Greek news summarization model trained on GreekSum.
This model is part of a series of models trained as part of our research paper: Giarelis, N., Mastrokostas, C., & Karacapilidis, N. (2024) GreekT5: Sequence-to-Sequence Models for Greek News Summarization [arxiv] The proposed models were trained and evaluated on the same dataset against GreekBART. For more information see the evaluation section below.

Training dataset

The training dataset of GreekT5-umt5-base-greeksum is GreekSum, which is the first news summarization dataset for the Greek Language.
This dataset contains ~151,000 news articles collected from News24/7, belonging to various topics (i.e., society, politics, economy, culture or world news).
For more information see: https://arxiv.org/abs/2304.00869

Training configuration

We trained google/umt5-base [580 million parameters (~2.37 GB)] on the GreekSUM train split using the following parameters:

• GPU batch size = 1
• Total training epochs = 10
• AdamW optimizer (e = 1e−8, β1 = 0.9 and β2 = 0.0999)
• Learning rate = 3e−4
• No warmup steps
• 32-bit floating precision
• Tokenization
  • maximum input token length = 1024
  • maximum output token length = 128
  • padding = ‘max_length’
  • truncation = True

Note: T5-based models use a multi-task architecture, the prefix ‘summarize: ’ was prepended in each training sample.

Evaluation

Approach|ROUGE-1|ROUGE-2|ROUGE-L|BERTScore ------------|-----------|-----------|-----------|------------- TextRank|18.10|5.76|13.84|68.39 GreekT5 (mt5-small)|14.84|1.68|12.39|72.96 GreekT5 (umt5-small)|25.49|12.03|21.32|72.86 GreekT5 (umt5-base)|**

Your obligations as a GreekT5 umt5 base greeksum deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Languages

Library