end-of-text marker at document breaks
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 4 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron reacts to the hidden marker the model uses to show that one document has ended and another begins. It is mostly silent otherwise. Turning it up did not produce a clear, consistent change in what the model wrote.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
, 2012<|endoftext|>Some weather conditions are simply not fun to ride in. As cyclists, windy days
-
0240.<|endoftext|>By 2003, plenty of data suggested that pesticide drift was harming
-
,0<|endoftext|>Guidelines for Restoring Cultural Landscapes⏎Identifying, retaining and preserving
-
rounded to the nearest integer.<|endoftext|>Christine Duncan, MD⏎Sickle cell disease (SCD) affects as
-
go<|endoftext|>A running scroll of past classes I have offered……..⏎DEFINITION: The
-
718<|endoftext|>Women who have kids later on in life may live longer, according to the findings
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: the user's name is not the same as the letter, and the user's name is not the same as the letter. Let's see the image (or not) (the image) or not the image. Let's see the image
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
position of formal authority. Alternatively then, if someone without a position of formal leadership is attempting to dominate the conversation, other team members may be less likely to acquiesce.⏎The above hypotheses were tested by Tost, G
-
is, in the United States.⏎The national Constitution enumerates the powers which may be exercised by the national government, and reserves all other powers “to the States respectively, or to the people.” Because of this dual or double character of our system
-
1x1, 2x2, 3x3, skill games) and 5x5 play, trying not<|endoftext|>Botanical Name: Asclepias exaltata⏎Plant Family: Milkweed family (Asclepi
-
transferred to Punjab;⏎iv) demand relating to halting the diversion of Punjab's river water.⏎There was a growing frustration as the Centre was not ready to acquiesce to their political demands. This provided a route to the dharma yud
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L3:236 excites
- L3:421 weakly suppresses
- L3:50 weakly excites
- L2:75 weakly suppresses
- L2:2328 weakly suppresses
- L3:189 weakly excites
- L1:2005 weakly excites
- L1:951 weakly excites
- L3:464 weakly excites
- L3:474 weakly excites
- L2:1920 weakly suppresses
- L1:1606 weakly excites
- L3:2850 weakly suppresses
- L3:2140 weakly suppresses
- L3:456 weakly suppresses
- L3:134 weakly suppresses
- L3:186 weakly suppresses
- L3:832 weakly suppresses
- L3:1594 weakly excites
- L2:1494 weakly suppresses
- L3:262 weakly suppresses
- L2:2788 weakly suppresses
- L1:1438 weakly excites
- L2:112 weakly excites
Feeds later layers
- L5:471 excites
- L5:11 excites
- L8:1365 excites
- L6:49 suppresses
- L9:81 excites
- L6:730 excites
- L18:2003 excites end-of-document separator token
- L10:507 excites
- L17:2580 weakly suppresses
- L10:227 weakly suppresses middle of fixed names and phrases
- L10:2957 weakly suppresses
- L8:1186 weakly excites
- L17:1751 weakly suppresses
- L5:1196 weakly suppresses
- L15:794 weakly suppresses
- L18:860 weakly excites
- L5:315 weakly excites
- L16:1709 weakly suppresses
- L7:105 weakly excites
- L13:2264 weakly excites
- L11:2341 weakly suppresses
- L9:1048 weakly suppresses
- L9:2880 weakly excites
- L13:283 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- hydration
- kickoff
- modeling
- lient
- PRESSION
- pects
- orian
- illac
- signature
- implications
Suppresses
- owered
- rays
- blick
- viz
- ưỡ
- ruta
- atient
- .TRUE
- FXML
- дал