distilbert-base-dutch-toxic-comments

Model description:

This model was created with the purpose to detect toxic or potentially harmful comments.

For this model, we finetuned a multilingual distilbert model distilbert-base-multilingual-cased on the translated Jigsaw Toxicity dataset.

The original dataset was translated using the appropriate MariantMT model.

The model was trained for 2 epochs, on 90% of the dataset, with the following arguments:

training_args = TrainingArguments(
    learning_rate=3e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    gradient_accumulation_steps=4,
    load_best_model_at_end=True,
    metric_for_best_model="recall",
    epochs=2,
    evaluation_strategy="steps",
    save_strategy="steps",
    save_total_limit=10,
    logging_steps=100,
    eval_steps=250,
    save_steps=250,
    weight_decay=0.001,
    report_to="wandb")

Model Performance:

Model evaluation was done on 1/10th of the dataset, which served as the test dataset.

Dataset:

Unfortunately we cannot open-source the dataset, since we are bound by the underlying Jigsaw license.