document start 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 hidden end-of-text marker that separates documents, which appears at the start of a new passage. Its effects on the model's writing were not clearly readable from the steering results.
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.
-
<|endoftext|>House in Space, from a NASA Remix Challenge. Credit: Cookieater20
-
<|endoftext|>10 Tips for Parents of Prospective College Students⏎1. Choosing a Career
-
<|endoftext|>Active transportation refers to non-motorized transportation modes, such as bicycling and
-
<|endoftext|>Public holidays in Pakistan are different than bank holidays in the country. While all of
-
<|endoftext|>Philosophy East and West 46 (2):143-
-
<|endoftext|>Cancer & Tumorigenesis⏎Cancer describes the disease with un
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 been thinking about about the things of the things of the things of the things of the things of the things of the things of the things of the things of the things of the things of the things of the things of the things of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
1 LED per segment and a decimal point.⏎Here is the pinout of a fairly typical display, you should always refer to the datas
-
government has taken no chances, shutting down parts of the city for rehearsals and tightening security as it readies for a show of might and pageantry.⏎Here is what the anniversary is about and what will happen.⏎Rebirth from civil war⏎
-
and sustainable – it would fall apart. Just as women enjoy a set of rights upon their husbands, husbands also possess certain rights over their wives in return.⏎Here is a look at some of the rights possessed by a husband in the realm of Islam
-
, I would ensure that there is a collaborative element; either have students work in partners working towards the same class puzzle.⏎Here is a link to a few puzzle ideas put out there by Quora members.⏎Pr
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
- L0:1301 suppresses
- L0:1224 suppresses
- L0:178 suppresses
- L0:84 suppresses
- L0:965 weakly excites
- L0:1279 weakly excites
- L0:220 weakly suppresses
- L0:508 weakly suppresses
- L0:305 weakly excites
- L0:632 weakly suppresses
- L0:2301 weakly excites
- L0:1680 weakly suppresses
- L0:188 weakly excites
- L0:326 weakly excites
- L0:2368 weakly suppresses
- L0:2951 weakly suppresses
- L0:2790 weakly suppresses
- L0:2347 weakly excites
- L0:1640 weakly excites
- L0:172 weakly excites
- L0:340 weakly excites
- L0:2657 weakly suppresses
- L0:2153 weakly suppresses
- L0:2442 weakly excites
Feeds later layers
- L2:325 excites
- L5:2680 suppresses
- L2:1457 excites
- L3:104 suppresses
- L2:93 suppresses
- L2:645 excites
- L3:110 excites
- L2:2160 excites
- L2:1173 suppresses
- L6:2293 excites 'in' or 'the' in 'up in the air'
- L3:537 excites
- L5:120 weakly suppresses
- L6:819 weakly suppresses
- L11:1309 weakly excites
- L9:584 weakly excites
- L4:178 weakly excites
- L10:3004 weakly suppresses
- L10:714 weakly excites
- L3:74 weakly suppresses
- L5:70 weakly excites
- L2:783 weakly excites
- L3:309 weakly excites
- L4:634 weakly excites
- L2:1885 weakly suppresses
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
- yna
- 筛
- acus
- 窨
- nâ
- yll
- 仿真
- สยาม
- distraction
- нт
Suppresses
- origin
- �
- راجع
- ecessary
- ithub
- 各项工作
- о
- MEDIATEK
- rect
- Origin