I wanted to see if the network would learn how to censor effectively or if it would try to fill in the censored parts. These pose a challenge to the LSTM, similar to how Google BERT used masked input in order to train their model. Articles handle these in two ways, either by using black bars, e.g. Some articles have redacted / censored parts.I wanted to see if the LSTM could replicate this format. There is an item number, object class (how dangerous it is), special containment procedures, which includes the dimensions and building materials of containment facilities, and a description of the object. These entities all exist within the same fictional universe and often reference one another, as well as different aspects of SCP lore. Each article in the training data represents a different supernatural entity. ![]() The SCP Foundation is a fictional organization which protects the world from supernatural threats. You can train a new model, continue training an existing model, or test a model to see what it outputs. Type "python lstm.py -h" for a list of command line arguments and description of each. It comes with a pretrained model, training data (~5,000 articles, 30M characters), and a scraper to show how the data was obtained. ![]() This program uses a Long Short Term Memory network to generate text in the style of SCP Foundation articles.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |