RO-Offense

This model is a fine-tuned version of readerbench/RoBERT-base on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.8411
• Accuracy: 0.8232
• Precision: 0.8235
• Recall: 0.8210
• F1 Macro: 0.8207
• F1 Micro: 0.8232
• F1 Weighted: 0.8210

Output labels:

• LABEL_0 = No offensive language
• LABEL_1 = Profanity (no directed insults)
• LABEL_2 = Insults (directed offensive language, lower level of offensiveness)
• LABEL_3 = Abuse (directed hate speech, racial slurs, sexist speech, threat with violence, death wishes, ..)

Model description

Finetuned Romanian BERT model for offensive classification.

Trained on the RO-Offense Dataset

Intended uses & limitations

Offensive and Hate speech detection for Romanian Language

Training and evaluation data

Trained on the train split of RO-Offense Dataset

Evaluated on the test split of RO-Offense Dataset

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4e-05
• train_batch_size: 64
• eval_batch_size: 128
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.2
• num_epochs: 10 (Early stop epoch 7, best epoch 4)

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Precision	Recall	F1 Macro	F1 Micro	F1 Weighted
No log	1.0	125	0.7789	0.7037	0.6825	0.7000	0.6873	0.7037	0.7132
No log	2.0	250	0.5170	0.8006	0.8066	0.8016	0.7986	0.8006	0.7971
No log	3.0	375	0.5139	0.8096	0.8168	0.8237	0.8120	0.8096	0.8047
0.6074	4.0	500	0.6180	0.8247	0.8251	0.8187	0.8210	0.8247	0.8233
0.6074	5.0	625	0.7311	0.8096	0.8071	0.8085	0.8064	0.8096	0.8071
0.6074	6.0	750	0.8365	0.8101	0.8117	0.8191	0.8105	0.8101	0.8051
0.6074	7.0	875	0.8411	0.8232	0.8235	0.8210	0.8207	0.8232	0.8210

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1+cu118
• Datasets 2.14.3
• Tokenizers 0.13.3