commas in dates and number lists
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. 22 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 lights up on commas that separate parts of a date, a year, or a list of numbers, such as 'November 18, 1918' or 'levels 1,2, and 3'. Turning it up made the model's writing turn into scrambled, repetitive text, so the effect is not easy to read.
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.
-
pass Level 2, they have to say level 1 and 2 correctly. To pass level 3, they have to say level 1,2, and 3 correctly. To pass level 4, the must say all of
-
20|⏎|Order Information:||Web: http://www.tandfonline.com/pricing/journal/VECE20|⏎When requesting a correction, please mention this item's handle: RePEc:taf:jeduce
-
ovember 5, 1818 – January 11, 1893
-
100⏎tspan=[t0 tend]; %need to use it in this form in MATLAB 5.⏎[t,y]=ode45('rhs',tspan,y0)%call the routine ode45⏎plot(t,y)
-
The basic⏎source of all minerals is the hot magma in the⏎interior of the earth.⏎(more content follows the advertisement below) A D V E R T I S E M E N T⏎When magma cools,⏎
-
THE⏎For more recent information about the government, see Facts about Albania.⏎Source: U.S. Library of Congress<|endoftext|>» 2010.aug.25 » We have the Russians to thank for the Internet. In 1
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 more than
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, and we know that,⏎Lines with the same slope are parallel.⏎Hence, the given lines are parallel.⏎Let's consider another set of equations.⏎y = 2x +3 and 4y = –2x+10
-
line, then they are perpendicular.⏎Let's see how to identify parallel and perpendicular lines.⏎Let's take some examples to understand this.⏎Considering two equations:⏎y = 2x +3 and y = 2x+5⏎On comparing
-
bigger than a two quarter⏎one quarter = two eighths therefore, three eighths is larger than one quarter⏎One quarter is 1/4 = 4/16 and three sixteenths is 3/1
-
, 24 = 4 ⋅ 6. So to solve the second part, on
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
- L21:164 suppresses
- L21:214 excites
- L14:35 suppresses
- L20:2320 suppresses
- L20:1298 suppresses
- L21:173 excites
- L19:1207 weakly suppresses
- L4:2864 weakly suppresses
- L19:534 weakly excites
- L19:726 weakly suppresses comma inside large numbers
- L18:2740 weakly suppresses
- L17:2940 weakly suppresses
- L5:539 weakly excites
- L17:118 weakly excites comma after a clause ending
- L18:115 weakly suppresses
- L19:220 weakly excites question mark ending a web question heading
- L14:309 weakly suppresses
- L20:966 weakly excites
- L17:215 weakly suppresses
- L19:1697 weakly suppresses
- L18:946 weakly suppresses
- L10:335 weakly excites commas in catalog, date, and list metadata
- L17:132 weakly suppresses
- L13:2204 weakly excites
Feeds later layers
- L24:51 strongly excites
- L26:17 strongly excites
- L23:239 strongly excites
- L25:222 excites
- L25:304 excites
- L25:146 excites
- L27:71 suppresses
- L23:262 excites
- L26:247 excites
- L25:109 suppresses
- L24:141 excites
- L26:103 excites blank space before a number
- L27:431 excites
- L26:224 excites
- L26:263 excites
- L25:217 suppresses
- L27:323 excites
- L27:513 excites
- L25:239 excites
- L27:514 excites
- L27:11 suppresses unclear: mixed fragments in lists and pronunciations
- L26:240 excites
- L27:220 excites end-of-text token between documents
- L26:199 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
- cre
- gl
- no
- standing
- She
- yield
- ine
- 결
- cum
- Indeed
Suppresses
- selected
- ;display
- "','
- !
- "','"
- !");⏎
- ','=',
- LSB
- !!
- powered