gap right before a number
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. 25 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 spot just before a number is expected, such as after 'over the age of' or 'kids under', where the next word will be a figure. Turning it up made the text repetitive and unfocused. It is weaker and less clear on the opposite side.
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.
-
in the two collaborating project laboratories. This work is<|endoftext|>“Holy Week: In Like a Lion, Out Like a Lamb” (Matthew 21:1-11; 27:11-66)⏎“In like a
-
23 types of pneumococcal bacteria that can lead to pneumonia, blood-stream infections and meningitis. A shot is recommended for everyone over the age of 65 and for those under 65 who have the same risk
-
bacteria. The disease can⏎affect the heart, joints, skin and brain.⏎Children in the age⏎group of 5 to 15 years are easy prey to this disease. The disease develops
-
project is eligible for School of Mathematics Doctoral Training Grant funding - please contact us for more information.<|endoftext|>História (São Paulo)⏎On-line version ISSN 1980-4369⏎DOMINGUES, Beatriz
-
Richeson, and such an appliance is utilized to expand that arch. The palatal expander is best used while a child is still growing, usually between ages 8 and 10.⏎"By age 7, a child will have
-
.⏎Yüksek Lisans Öğrenimine MSÜ Barbaros DEBİM Enstitüsünde devam etmekte aynı zamanda enerji sistem ve modelleme üzerine araştırmalar yapmaktadır.<|endoftext|>-
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 key to a room even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even roomy even
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ched, soft rays. The anal fin has one spiny ray and 24-26 soft rays. The caudal fin is truncated.⏎The upper part of the body is yellowish or brownish, the lower part is white
-
zoom in or out). Your responses should demonstrate critical thinking. Provide justification for your analysis:⏎How many earthquakes are listed on this map?⏎How many earthquakes are over 6.0 over the week? How many between 4.5 and
-
(Hindi:कामायनी) (1936) is a Hindi epic poem (Mahakavya) by Jaishankar Prasad (1889–1937). It is
-
question for which a “yes/no” is perfectly adequate. This is not the case with a feature that has a specific tolerance i.e.: a nominal and an upper and lower spec limit (1.0 +/- .005). It might
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
- L20:54 excites
- L20:3033 suppresses
- L19:305 weakly excites
- L22:349 weakly excites
- L3:1179 weakly excites
- L0:178 weakly suppresses
- L18:251 weakly excites
- L18:125 weakly suppresses
- L20:541 weakly excites
- L19:859 weakly excites
- L0:1301 weakly suppresses
- L0:1224 weakly suppresses
- L15:184 weakly excites
- L17:292 weakly excites
- L3:283 weakly excites
- L18:178 weakly excites
- L0:2792 weakly suppresses middle pieces of split words
- L20:270 weakly excites
- L17:17 weakly suppresses
- L14:1289 weakly excites
- L24:685 weakly suppresses
- L7:92 weakly excites
- L14:20 weakly excites
- L11:844 weakly excites
Feeds later layers
- L27:1164 weakly excites
- L26:1482 weakly excites
- L27:1985 weakly excites
- L26:1952 weakly excites
- L26:1029 weakly excites
- L26:637 weakly excites
- L27:2381 weakly excites
- L27:1846 weakly excites
- L27:848 weakly excites
- L26:1213 weakly excites
- L27:2131 weakly suppresses
- L27:1690 weakly excites
- L26:353 weakly suppresses
- L26:2969 weakly excites
- L27:285 weakly suppresses
- L27:837 weakly excites
- L27:1410 weakly excites "as well" and "as much as"
- L27:14 weakly excites
- L26:2865 weakly excites
- L27:2218 weakly excites
- L27:893 weakly suppresses
- L26:1509 weakly suppresses
- L27:2234 weakly suppresses
- L26:1356 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
- ise
- 娩
- wort
- unta
- gne
- êt
- azine
- ahoo
- ServerError
- wj
Suppresses
- setDefault
- cí
- pery
- ума
- os
- even
- =os
- chimp
- -DD
- hers