Measure the baseline
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.
MODEL INTERPRETABILITY / EXPERIMENT 01
Who does Inkling
think it is?
One prompt. One neuron pushed up or down. Explore how single-neuron interventions change the way Inkling-Small describes itself.
SUBJECT MODELInkling-Small
DECODER LAYERS42
TOTAL PARAMETERS276B
ACTIVE PARAMETERS12B
01 / OBSERVATIONS
Persona library
Search the replies, compare push directions, or look up a neuron by its coordinates.
The first run is pending
Import a run from the experiment runner to browse actual Inkling-Small replies.
02 / ARCHITECTURE
Layer atlas
All 42 decoder layers, from the two dense blocks to the mixture-of-experts backbone.
Import a run to see its coverage.
Dense Mixture of experts Observed
03 / PROTOCOL
How a neuron is pushed
A text-only adaptation of the original Neuron Persona Library experiment.
Change one direction
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‖
Read the reply
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?
Layer and neuron identify a dense coordinate. Mixture-of-experts coordinates also include expert type and index. The same residual intervention uses that expert’s down-projection column directly. For routed experts, this deliberately bypasses routing and tests the direction regardless of which experts the token selected.
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.