xlm-roberta-base-toxicity (Spanish)

This model is a fine-tuned version of xlm-roberta-base on 2 datasets, labelled with not_toxic (0) / toxic (1) content from news or tweets.

• a private one, provided by @Newtral, containing both tweets and news.
• one used for data augmentation purposes, containing only news, obtained from SurgeHQ.ai

This model can not be used for commercial purposes

Training and evaluation data

The test dataset was provided by @Newtral and was kept fixed.

It achieves the following results on the evaluation set:

• eval_loss: 0.4852
• eval_f1: 0.8009
• eval_accuracy: 0.901
• eval_runtime: 13.6483
• eval_samples_per_second: 366.347
• eval_steps_per_second: 22.933
• epoch: 5.0
• step: 3595

Training procedure

• Cleaning
• Data Augmentation
• Optuna for Grid Search
• Shap for interpretability

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.889038893287002e-06
• train_batch_size: 16
• eval_batch_size: 16
• seed: 37
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Framework versions

• Transformers 4.18.0
• Pytorch 1.10.2+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1