end-of-document token boundary
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. 17 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 fires most strongly on the marker that signals the end of one web document and the start of another. It goes quiet, or negative, on dashes and colons inside citation or metadata lines. Turning it up did not give a clear shared theme in the text it 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.
-
liu.<|endoftext|>Bond Price Calculator⏎With this bond price calculator, we aim to help you calculate
-
stricting blood vessels?<|endoftext|>There is growing recognition that a sedentary lifestyle is being driven, at least
-
,0<|endoftext|>Guidelines for Restoring Cultural Landscapes⏎Identifying, retaining and preserving
-
, 2012<|endoftext|>Some weather conditions are simply not fun to ride in. As cyclists, windy days
-
Influence Insider?<|endoftext|>As the world scrambles to find antiviral treatments and vaccines for the novel
-
to review it. We're adding the slope at a single point to the slope over a big area, which is something I can't quite work out visually.<|endoftext|>The world produces more than enough food to feed everyone, yet each night more than
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: I have a 300-100-100-100-100-100-100-1-10-1-1-1-1-1-1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
backgrounds.⏎|Original language||English (US)|⏎|Number of pages||12|⏎|Journal||Classical and Quantum Gravity|⏎|State||Published - Jul 7 2003|⏎ASJC Scopus subject areas⏎
-
ENNEDY, Chi Kin John LEE|⏎|Place of Publication||Abingdon, Oxon|⏎|Publication status||E-pub ahead of print - Aug 2020|<|endoftext|>10. Rebinding an Old Book
-
and A[1:0] for ARM instruction fetches (nOPC=0 and MAS[1:0]=10).⏎Article last edited on: 2008-09-09 15:4
-
Jiangsu 211171, China⏎*Authors to whom correspondence should be addressed.⏎Received: 2020-8-4 / Accepted: 2020-9-8 / Published: 2
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
- L8:4 suppresses
- L8:13 excites
- L8:7 suppresses
- L11:2 excites unclear: blank before a year or number
- L8:16 excites
- L6:13 suppresses
- L8:10 suppresses
- L6:12 suppresses
- L8:3 suppresses
- L9:5 suppresses
- L3:464 suppresses
- L7:64 excites
- L7:191 suppresses
- L6:7 excites
- L5:50 suppresses
- L8:14 suppresses
- L7:102 excites
- L6:18 suppresses
- L8:12 suppresses
- L10:46 suppresses
- L13:129 suppresses
- L5:158 excites
- L15:5 suppresses
- L7:122 suppresses
Feeds later layers
- L27:725 weakly suppresses
- L27:84 weakly excites
- L27:1665 weakly excites
- L27:1974 weakly excites
- L27:2211 weakly excites
- L27:314 weakly excites
- L27:1995 weakly excites
- L27:2116 weakly excites unclear: first and boundary tokens
- L27:82 weakly excites
- L27:2326 weakly excites
- L27:107 weakly excites
- L19:423 weakly excites words that set up a following clause
- L27:194 weakly excites
- L19:669 weakly excites
- L27:1371 weakly excites
- L27:1 weakly suppresses
- L27:47 weakly excites
- L27:823 weakly suppresses
- L27:37 weakly excites
- L27:1345 weakly suppresses
- L27:301 weakly excites
- L27:1466 weakly excites
- L27:2110 weakly suppresses
- L27:801 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
- ilian
- otional
- uggage
- usal
- ikers
- Wants
- ughters
- ailer
- Board
- htags
Suppresses
- 举
- правило
- именно
- recourse
- 打击
- —it
- assortment
- <source
- plaintext
- нос