This model is a fine-tuned version of Google's bert-base-uncased that classifies a given artifact description into one or multiple categories of bias: subjective, jargon social, gender. The model achieves an accuracy of 83% given biased descriptions.

The dataset used to fine-tune the model is Michael C. Carlos Museum's internal collections database. See our paper for more details on the partnership, model, and pipeline. The model's input token limit is 512 tokens; the same as the original BERT model. See our github repository for our compelete solution.