Model Card for gplus

This model is a facebook/bart-large fine-tuned on non-toxic comments from jigsaw_toxicity_pred dataset. Only a subset (20%) of the non-toxic comments were used for training this dataset.

Model Details

This model is not intended to be used for plain inference, even though it is unlikely to predict toxic content. It is intended to be used as "utility model" for detecting and fixing toxic content as its token probability distributions will likely differ from comparable models trained/fine-tuned over toxic data. Its name gplus refers to the G+ model in Detoxifying Text with MARCO: Controllable Revision with Experts and Anti-Experts.

Model Description

• Developed by: [tteofili]
• Shared by : [tteofili]
• License: [apache-2.0]
• Finetuned from model : facebook/bart-large

  Bias, Risks, and Limitations

  This model is fine-tuned over non-toxic comments from jigsaw_toxicity_pred, it is unlikely to produce toxic content. Nevertheless, this model should only be used in combination with other models for the sake of detecting / fixing toxic content, see for example Detoxifying Text with MARCO: Controllable Revision with Experts and Anti-Experts.

  Evaluation

  This section describes the evaluation protocols and provides the results.

  Testing Data, Factors & Metrics

  Testing Data

  This model was tested on jigsaw_toxic_pred testset.

  Metrics

  Model was evaluated using perplexity (on the MLM task).

  Results

  Perplexity: 1.02