Ask “Who are you?” using Inkling-Small’s chat template with reasoning disabled. Record the greedy reply and the median norm of the layer’s residual vectors across prompt tokens, excluding the first token.
The Inkomplete Inkling Persona Atlas
This atlas adapts the single-neuron steering method from Joshua Fonseca Rivera’s Neuron Persona Library. Each test pushes one neuron up or down and asks Inkling-Small “Who are you?”
It’s Inkomplete because I only used the credits on Modal’s free plan. I wanted to see how far I could get without spending anything.
SUBJECT MODELInkling-Small
DECODER LAYERS42
TOTAL PARAMETERS276B
ACTIVE PARAMETERS12B
Persona library
The first run is pending
Import a run from the experiment runner to browse actual Inkling-Small replies.
Layer atlas
Import a run to see its coverage.
Dense Mixture of experts Observed
How a neuron is pushed
Add or subtract a normalized down-projection column after the chosen decoder block. Scale the push by the baseline residual norm and the recorded strength.
h′ = h ± α · r · w / ‖w‖
Compare each steered reply with the unsteered baseline. Optional labels and scores live alongside the original text. Each observation describes one prompt and intervention.
How do expert neurons work here?
“L16 · Routed 128” means we pushed a neuron from routed expert 128 in layer 16. Layers 0 and 1 are dense and have no experts. In the other layers, both shared experts run on every token, and the router picks six routed experts per token. We add the chosen neuron’s direction after the layer runs, so the label doesn’t tell you which experts the router picked.
What does a persona label mean?
A persona label describes a reply in which the model took on a different identity. Labels are interpretations of individual replies. Other prompts, checkpoints and intervention settings can yield different replies.
Where do the data come from?
The experiment runner exports the checkpoint, prompt, baseline and intervention replies as JSON. This viewer loads run files in the current browser tab. Model architecture comes from the published Inkling-Small configuration.