period after month or initial
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 full stop that follows a short abbreviation, like the dot in 'Aug.' or after a person's initial such as 'F. Scott'. It is quiet on the full stop that closes a quotation or bracket. Turning it up or down gave text that did not show a clear link to this pattern.
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.
-
making glass and an ever-expanding range of uses, American-made glass has become a great source of interest for historical researchers and collectors alike.⏎On Saturday, Jan. 10, at the Coventry Community Center,
-
ID also called on Chile to speed up its judicial process and create a national plan to search for disappeared persons.⏎From Jurist, Aug. 23. Used with permission.<|endoftext|>Oise River, river that rises
-
that pheromones – in contrast to chemical pesticides – have no adverse effects on animals or humans.<|endoftext|>F. Scott Fitzgerald⏎Francis Scott Key Fitzgerald (September 24, 1
-
to perform basic statistical functions using both a scientific calculator and Microsoft Excel.⏎Saturday, Aug. 23, 2-4:30 p.m.⏎Basic Statistics
-
person, except when referring to Hana, and without using quotation marks. Since there are no quotation marks, Ondaatje is using language as<|endoftext|>A.N.L. - folds (1996), for computer generated tape is a
-
through our internet radio service in Model X. Because AM audio quality can be very poor, we offer internet radio to give<|endoftext|>A New Brunswick Ghost Story⏎S. E. Schlosser⏎It was a cold winter and the snow was deep
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: 2.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
control of bleeding following mild injury while preventing inappropriate intravascular thrombosis. (See "Overview of hemostasis".)<|endoftext|>How to save your life in a fire or other emergency⏎For those
-
light(s). It turns out, perhaps somewhat surprisingly, that there is a simple and intuitive
-
War on Terror’. Curiously, despite their proliferation, they have received relatively little academic attention. In
-
in hard copy. Check the Northern Research Station web site to request a printed copy of this publication.⏎- Our on-line publications are scanned and captured using Adobe Acrobat. During the capture process some typographical errors may occur. Please contact Sharon Hobrl
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
- L4:7 strongly excites
- L2:9 strongly suppresses sentence-final periods vs dots in code
- L15:313 strongly excites decimal point inside numbers
- L5:168 suppresses
- L14:422 excites periods after initials and abbreviations
- L11:132 excites period ending an opening topic sentence
- L1:130 suppresses
- L3:595 excites
- L5:155 suppresses
- L8:143 suppresses
- L0:36 excites
- L5:1304 excites
- L15:241 excites
- L4:295 excites
- L15:123 weakly suppresses
- L0:348 weakly excites space token right after a period
- L14:51 weakly excites colon that introduces a statement
- L5:574 weakly excites
- L3:1555 weakly suppresses
- L8:16 weakly suppresses
- L6:305 weakly suppresses
- L9:16 weakly suppresses unclear: periods inside numbers and names
- L0:2838 weakly suppresses
- L0:3046 weakly excites
Feeds later layers
- L17:642 strongly suppresses
- L19:831 strongly suppresses
- L17:1218 strongly excites
- L18:713 strongly suppresses
- L20:505 strongly suppresses
- L17:347 strongly suppresses full stops ending informative sentences
- L18:1437 strongly suppresses
- L18:791 strongly suppresses
- L17:2084 excites
- L17:1772 suppresses
- L20:2301 suppresses
- L19:344 suppresses periods ending explanations
- L17:990 suppresses
- L21:1086 suppresses end of sentence before a claim or statistic
- L20:962 suppresses decimal points and dots in numbers
- L17:1155 suppresses opening parenthesis after a sentence
- L23:179 suppresses
- L27:145 suppresses dots inside web addresses
- L19:902 excites
- L24:353 suppresses decimal point inside a number
- L22:67 suppresses
- L19:2497 suppresses
- L22:282 suppresses period then line break in lists
- L21:524 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
- alte
- inary
- ENTITY
- olated
- 椰
- =com
- AEA
- Ζ
- ват
- ungal
Suppresses
- plier
- гин
- 一个
- 打法
- ento
- ään
- しかも
- At
- ора
- -note