end-of-document boundary token
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. 16 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 on the invisible marker the training data uses to separate one web page from the next, which appears in the middle of a sentence or list where a new document starts. Turning it up made the model repeat newlines and dots, so it seems tied to where a text stops and another begins.
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.
-
994- Drexel University. All rights reserved.⏎The Math Forum is a research and educational enterprise of the Drexel University School of Education.<|endoftext|>Maybe you’ve tried intermittent fasting (IF) to shed a few (pand
-
bonsai tree is the most common reason for its decline.⏎Mary Lougee has been writing for over 10 years. She holds a Bachelor's Degree with<|endoftext|>- freely available⏎Sustainability 2013, 5(7
-
VS is a promising alternative strategy to produce the functional and thermostable reverse transcriptase.⏎All Science Journal Classification (ASJC) codes⏎- Insect Science<|endoftext|>An account of the religion of
-
the China site (in Chinese or English) for best site performance. Other MathWorks country sites are not optimized for visits from your location.⏎Contact your local office<|endoftext|>It’s PrAACtically Martin Luther King Jr Day⏎Looking for activities and
-
the China site (in Chinese or English) for best site performance. Other MathWorks country sites are not optimized for visits from your location.⏎Contact your local office<|endoftext|>The point in the Moon's orbit where it lies diametrically⏎opposite
-
S||51.00W To 54.00W|⏎|rs_class_santarem.tif||163.4 MB||<|endoftext|>English-to-Spanish medical translation will be more important than ever in the United
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: (for example, for example, in the same or or or or or or or or or or (for example, for example) and for the same or or or (or or or 1-1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of allergens and irritants, but it can eventually lead to structural damage in your home if left untreated. A dehumidifier or air conditioner can help.⏎Answer: True. Radon—a colorless, tasteless and odorless gas
-
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document⏎Unformatted text preview: itself, and if ( ) ( ) lim lim x
-
To access and follow an EdPlace learning plan, please download the attachment at the end of the article.⏎How can I use them to support my child?⏎There are several ways in which
-
shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally blurred sections. Sign up to view the full version.View Full Document⏎Unformatted
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
- L15:81 strongly suppresses line breaks inside running text
- L14:56 suppresses document ends and list line breaks
- L10:84 suppresses
- L13:10 excites
- L10:259 excites
- L8:32 excites
- L9:25 excites
- L1:1 excites
- L15:123 suppresses
- L14:1067 suppresses
- L15:82 suppresses
- L12:5 excites
- L14:51 excites colon that introduces a statement
- L9:7 excites
- L15:44 suppresses
- L10:306 suppresses
- L1:1279 suppresses
- L9:2729 suppresses
- L9:605 excites
- L6:216 excites
- L15:385 excites
- L9:12 weakly suppresses
- L0:37 weakly excites
- L15:1620 weakly excites period ending a sentence in essays
Feeds later layers
- L17:1176 strongly suppresses
- L17:1953 strongly excites
- L17:1568 strongly excites
- L19:1479 strongly excites
- L18:1950 strongly suppresses opening parenthesis after a name or title
- L18:2202 strongly excites
- L19:441 excites
- L19:624 excites
- L18:2922 excites
- L20:1081 suppresses
- L20:1085 excites
- L21:130 excites
- L22:506 weakly excites
- L26:2463 weakly suppresses
- L17:537 weakly excites
- L20:467 weakly excites
- L17:2317 weakly excites
- L17:1024 weakly excites
- L25:1274 weakly suppresses
- L17:1616 weakly excites
- L27:1065 weakly suppresses
- L25:1740 weakly suppresses
- L19:1007 weakly excites
- L18:705 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
- 们都
- less
- 们的
- 四方
- icks
- Whenever
- 人都
- bölüm
- Интер
- partes
Suppresses
- enstein
- nung
- omen
- yne
- 亿吨
- �
- ughter
- etailed
- Hussein