Overview
SelfPortrait explores generating replies in a personal conversational style. I'm working on v2; the public repository covers the earlier version.
The earlier implementation
The pipeline processes conversation examples, builds a training dataset, fine-tunes a model, and runs a local inference server. A client puts a suggested reply in the text box for review before sending (I reserve the right to disagree with myself).
Even the preprocessing involves choices: where to split a conversation, when consecutive messages count as one turn, and which examples are useful for training. These choices affect the result before the model sees any data.