Embedded Files
2025
A high-level overview of fine tuning a model to contain my memories, personality, and speaking tone. All from an iPhone, using local compute (2 Mac Studios), to run on a 7b model (Mistral) that is capable of running inference on an iPhone.
I created the data pipelines and training recipes myself, learning on the fly!
The Digital Twin can be kept up to date by regularly fine-tuning on a user's data.
The final prototype built on Mistral 7b ran on less than 15GB of RAM. It is feasible, with further optimizations, to get this performance on, at most, a 1b model using 4GB of RAM or less. Small enough to query very fast on an iPhone.
This unlocks new possibilities such as user simulations, new personalization options, and more.
Google Sites
Report abuse