Release 0.17 ·
From a prompt to a trained model
The assistant builds real models from a prompt, sized for the machine it is running on and the dataset it is given.
In 0.17 the assistant stopped describing models and started building them: a prompt becomes blocks on the canvas, sized for the hardware NEURAX detected and the dataset it profiled.
Datasets are profiled where they are: on this machine or in a bucket. The assistant receives the profile: shapes, counts, balance. Never a path, and never the contents.
Training reports only real runs, and anything shown as an example is labelled as one. Decision now comes before Training, so the person who signs sees the numbers before anything runs.
Small things that mattered: a missing Python package is reported as something to install rather than a server error, and an empty canvas is no longer analysed.