HAIM - Towards Controllable Text Generation
As a step towards controllable text generation, we developed HAIM, a language model that can fill-in synthetic text between a human-written beginning and a human-written ending. HAIM allows the user to guide the text generation more accurately than its predecessors, by feeding the model with an ending and a desired length, in addition to a beginning. We’re releasing a demo of HAIM-Large, a variant of the model with 345M parameters trained on the 40GB OpenWebText dataset.
Give it a try!