metadata

language:
  - es
thumbnail: null
tags:
  - summarization
  - mt5
  - spanish
license: apache-2.0
datasets:
  - larazonpublico
  - es
metrics:
  - rouge

mt5-small-spanish-summarization

Model description

This is a mt5-small model finetuned for generating headlines from the body of the news in Spanish.

Training data

The model was trained with 58425 news extracted from the La Raz�n (31477) and P�blico (26948) newspapers. These news belong to the following categories: "Espa�a", "Cultura", "Econom�a", "Igualdad" and "Pol�tica".

Training procedure

It was trained with Google Colab's GPU Tesla P100-PCIE-16GB for 2 epochs.

Hyperparameters

{evaluation_strategy = "epoch", learning_rate = 2e-4, per_device_train_batch_size = 6, per_device_eval_batch_size = 6, weight_decay = 0.01, save_total_limi t= 3, num_train_epochs = 2, predict_with_generate = True, fp16 = False}

Eval results

metric	score
rouge1	44.03
rouge2	28.2900
rougeL	40.54
rougeLsum	40.5587

BibTeX entry and citation info

@inproceedings{ mt5lrpjosmunpen,
  year={2020},
  author = {Jos� Manuel Mu�iz Pe�a},
}