Bart from Scratch

A full implementation of the BART transformer written from scratch, without any pretrained weights. I trained it on the CNN/Daily Mail dataset for article summarization and added both greedy and beam search decoding. The architecture is configurable through a JSON file, and it ships as both a notebook and a structured codebase.