Model Card for roberta-base-formality

FacebookAI/roberta-base model, fine-tuned as a binary classifier of formality on the GYAFC dataset.

Model Details

Model Description

• Developed by: @cointegrated
• Model type: classifier
• Language(s) (NLP): English
• License: probably MIT (as the base model), but please double-check the GYAFC terms of usage

The model has been trained for 5 epochs on the GYAFC train split with batch size of 32 and Adam optimizer with learning rate of 1e-5.

The evaluation results are: 99.13% ROC AUC on the dev split, and 99.20% ROC AUC on the test split.

Uses

Classification of formality of texts; originally intended as a tool for evaluating text style transfer.

Class 1 denotes formal texts, and class 0, informal texts.

Note that this model is not expected to perform well in any languages other than English or in any domains distant from the GYAFC data distribution (Yahoo Answers sentences for "Entertainment & Music" and "Family & Relationships" categories). Also, as a binary classifier, this model may not handle intermediate degrees of formality inadequately. If you want to rank texts by varying degrees of formality, s-nlp/roberta-base-formality-ranker (a later version of this model, trained more contrastively) might be a better choice.