end-of-document marker 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. 1 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 invisible marker the model uses to mark where one web page or document ends and another begins. Turning it up made the model's writing wander into odd, unfocused directions, but the link to the marker is the clear part.
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.
-
caused<|endoftext|>This preview shows pages 1–2. Sign up to view the full content
-
he advocated for peace with the European Americans⏎Date de dernière mise à jour : 05/07/2021⏎Ajouter un commentaire<|endoftext|>By DANIEL SUDDEATH⏎NEW ALBANY —⏎Judging
-
asked<|endoftext|>Counties that gained Wal-Mart stores during the 1990s f
-
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
-
5mm or 120 roll film into a developing canister.⏎- Processing the film⏎Step-by-step instructions to develop your film.⏎Leave a Comment<|endoftext|>As human activities drive Earth’s rapidly changing climate, there is an urgent need to
-
a Chemistry Tutor<|endoftext|>Those who fight for us, also need people to fight for them.⏎The military
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 think I have been thinking that the most important thing is to have a strong foundation in both the subject and the skills that are used to have a strong foundation in both the subject and the skills that are used. That is the thing I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cardinal numbers. You will need to add the cardinal numbers together and the fractions together. Finally, you need to check whether you can simplify the answer.⏎Step 1 - Addition, but don't forget to simplify⏎Step 2 - Drag the
-
much more!⏎Discover the world of gorillas and other primates their remarkable intelligence and social behavior⏎66
-
today is totally different and not close to what used to be American music. The question which many people ask is the vents which led to such intensive change. This essay focuses on American music genre Jazz rap which has engulfed the original culture of America
-
attle to produce sound. They will also be able to hold their head steady.⏎- Step 7: Hug them at four to five months and they should try to hug
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
- L0:1224 suppresses
- L0:178 suppresses
- L0:1301 suppresses
- L0:462 weakly suppresses
- L0:1369 weakly suppresses
- L0:1075 weakly excites
- L0:1289 weakly suppresses
- L0:317 weakly excites
- L0:42 weakly excites
- L0:60 weakly excites
- L0:88 weakly excites
- L0:1046 weakly excites
- L0:1141 weakly excites
- L0:191 weakly suppresses
- L0:1126 weakly excites
- L0:1041 weakly excites
- L0:2515 weakly excites
- L0:1173 weakly suppresses
- L0:1885 weakly suppresses
- L0:748 weakly excites
- L0:402 weakly suppresses
- L0:1316 weakly suppresses
- L0:1554 weakly excites
- L0:2961 weakly suppresses
Feeds later layers
- L16:437 weakly excites
- L15:1058 weakly suppresses
- L8:1001 weakly excites
- L4:2799 weakly suppresses
- L23:929 weakly suppresses
- L17:1585 weakly suppresses
- L25:515 weakly suppresses
- L10:1095 weakly suppresses
- L17:2923 weakly excites
- L8:717 weakly suppresses
- L5:2683 weakly suppresses
- L20:829 weakly suppresses
- L24:1256 weakly excites
- L14:2505 weakly excites
- L5:2086 weakly suppresses
- L8:1007 weakly suppresses
- L15:1540 weakly excites
- L4:233 weakly excites line break in the middle of a sentence
- L8:34 weakly suppresses
- L21:1175 weakly suppresses
- L6:229 weakly suppresses
- L10:2880 weakly excites
- L15:107 weakly excites
- L16:726 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
- nost
- posable
- 'h
- .asList
- .t
- ptide
- 慑
- iband
- loc
- >t
Suppresses
- /cop
- DOG
- KeyType
- CUT
- editing
- -strokes
- コピー
- alin
- 又是
- areas