Unedited
Qwen3.5-0.8B as released, the labels in its prompt
Each request, routed.
hedwigaihedwigaihedwigai
Make a small model your router
Continue in
Give labels and a few examples of each. Qwen3.5-0.8B gets one new neuron per label in its last layer. The neurons are fitted by least squares from one forward pass per example, with no gradients. You get a scorecard on examples the fit never saw, and the neurons as a patch. The post explains how it works.
1. Pick a task
2. The scorecard
The scorecard appears here once the minion is shaped.
The examples are split three to one to one. The first part fits the neurons, the second picks their strength, and the third is scored here. With five examples a label, the third part holds one example a label, so read a small check set as a rough sign.
3. The examples
Loading the recorded tasks…
Sign in to run
These two tasks are free to look at; sign in to shape your own.
4. Compare
Minion
The same model with one neuron a label added