Training Lab
This page is a running home for training experiments, eval ideas, half-finished notes, and the older dictation work that led into the current voice memo stuff.
What this is
I'm using this route to keep the work in one place while I learn. Some of it is careful benchmark design, some of it is me trying things, getting them wrong, and writing down what changed.
Right now the main thread is voice memo extraction. The older dictation work is still here because it explains how I ended up caring about these smaller note-shaped tasks.
Stuff I'm Working On
Current thread
This is the newer thread. I'm trying to turn short voice memos into cleaner, more usable artifacts without pretending a tiny model knows more than it does.
Foundation
This is the older thread. It started with spoken commands and shell syntax, but it still shapes how I think about cleanup, normalization, and where models should stop and code should take over.
One Current Example
Need to talk to Maya about the deck or maybe just send it over first, I'm not sure which is less annoying.
{
"title": "Decide how to share deck with Maya",
"intent": "none",
"target": ""
}Recent Writing
Pulling titles and lightweight actions from voice memos, how to measure the result, and what to do when an extraction is useful but not ready to finalize.
A reader-first walkthrough of a new semantic eval for tiny local models, including what it measures, how the scores work, and what the first local runs show.
The practical design document behind core_eval_v2, from product truth and scoring layers to calibration rules and what still feels unfinished.
You can iterate forever on training data. At some point you have to ask whether the model is the right tool for the job.