multilingual_sentiment_newspaper_headlines

This model is a fine-tuned version of bert-base-multilingual-cased on a dataset of 30k newspaper headlines in German, Polish, English, Dutch and Spanish. The dataset contains 6k headlines in each of the five languages. The newspapers used are as follows:

• Polish: Fakt, Rzeczpospolita, Gazeta Wyborcza
• English: The Times, The Guardian, The Sun
• Dutch: De Telegraaf, NRC, Volkskrant
• Spanish: El Mundo, El Pais, ABC
• German: Suddeutsche Zeitung, De Welt, Bild

It achieves the following results on the evaluation set:

• Train Loss: 0.2886
• Train Sparse Categorical Accuracy: 0.8688
• Validation Loss: 1.0107
• Validation Sparse Categorical Accuracy: 0.6434
• Epoch: 4

import torch
from transformers import AutoTokenizer, TextClassificationPipeline,TFAutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("z-dickson/multilingual_sentiment_newspaper_headlines")
m1 = TFAutoModelForSequenceClassification.from_pretrained("z-dickson/multilingual_sentiment_newspaper_headlines", from_tf=True)
sentiment_classifier = TextClassificationPipeline(tokenizer=tokenizer, model=m1)

sentiment_classifier('Brazylia: Bolsonaro wci±ż nie uznał porażki. Jego zwolennicy blokuj± autostrady')
[{'label': 'negative, 0', 'score': 0.9989686012268066}]

Training hyperparameters

The following hyperparameters were used during training:

• optimizer: {'name': 'Adam', 'learning_rate': 2e-05, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-07, 'amsgrad': False}
• training_precision: float32

Training results

Train Loss	Train Sparse Categorical Accuracy	Validation Loss	Validation Sparse Categorical Accuracy	Epoch
0.8008	0.6130	0.7099	0.6558	0
0.6148	0.6973	0.7559	0.6200	1
0.4626	0.7690	0.8233	0.6368	2
0.3632	0.8229	0.9609	0.6454	3
0.2886	0.8688	1.0107	0.6434	4

Framework versions

• Transformers 4.26.0
• TensorFlow 2.9.2
• Tokenizers 0.13.2