sentence-ending full stops in web text
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. 9 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 full stops that end a sentence in web pages and articles, especially after a name or a title. Its negative side barely matters. Turning it up or down changed the topic of the output, but not in a clear way.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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.
-
-blame/failure/low self-esteem Sadness/depression: loss . In Building Self-Esteem in Children Who Are Adopted or Fostered, Dr. Sue offers simple and practical advice to those supporting children aged 7+ to help
-
yellow wire is burned and the wire on the right side is deep inside of the cavity of the oven, the spade prong was stuck on the old element. The connection at the right pull off with a pair of pliers and just needs
-
devices and you want to represent them in JAVA. Current states of a class’s co… In this chapter, we will look into the concepts - Classes and Objects. Java class objects exhibit the properties and behaviors defined by its class. Java Objects.
-
As an application developer, you are free to use any of these ports. If definitions are helpful to you, use these vocabulary terms in order to get you started. Administer digital networking port b10 worksheet b17. Apple macos
-
when you register 4.⏎Discussion questions for ‘The Death and Life of the Great Lakes’⏎Skip to search form Skip to main content You are currently offline. Some features of the site may not work correctly. They are home to 3
-
by zero. Even on our own wiki, at one point.⏎Handwriting is such serious business that some people actually consider it an art to make words on paper. Never mind what words they actually wrote, whether they used printing or cursive is
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 number of the phone. A number is a 10-digit number, which is a 10-digit number. It is a 10-digit number, which is a 10-digit number, which is a 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
⏎The main crack<|endoftext|>analysis of algoritms⏎robert.kern at gmail.com⏎Fri Sep 10 00:10:22 CEST 2010⏎On 9/9/10
-
17:55 EST⏎- ASU instrument on NASA rover helps identify outcrop of long-sought
-
that has been in use worldwide for over 200 years. It has been available on the<|endoftext|>Register New Player⏎Welcome to our world of fun trivia quizzes and quiz games:⏎The Battle of Badr⏎"The Battle of Badr was
-
initial batch of 41 country pages, with more to follow. They are:⏎Afghanistan, Angola, Argentina, Bahrain, Belarus, Belgium, Burkina Faso, Burundi, Cambodia, Cameroon, Colombia, Costa Rica, Czech
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
- L8:802 excites
- L8:625 weakly suppresses
- L7:1053 weakly suppresses
- L5:102 weakly excites
- L5:2422 weakly suppresses
- L8:324 weakly suppresses
- L7:58 weakly suppresses
- L7:1418 weakly excites
- L7:133 weakly suppresses line break inside a sentence
- L2:772 weakly excites
- L7:1040 weakly suppresses
- L2:1873 weakly excites
- L1:998 weakly excites
- L7:2976 weakly suppresses
- L3:2483 weakly excites
- L3:2973 weakly suppresses
- L5:2332 weakly suppresses
- L1:2877 weakly suppresses
- L3:2312 weakly excites
- L8:52 weakly excites
- L5:223 weakly excites
- L5:1026 weakly suppresses
- L4:2735 weakly suppresses
- L3:1790 weakly excites
Feeds later layers
- L18:1105 excites
- L15:1820 excites
- L14:2497 excites
- L13:1086 excites
- L15:2188 excites
- L15:1619 excites
- L10:258 excites
- L10:348 excites
- L15:1448 excites
- L15:1884 excites
- L12:195 excites
- L16:2634 weakly excites
- L10:97 weakly excites
- L15:2887 weakly excites
- L20:2082 weakly suppresses
- L21:530 weakly suppresses
- L12:1144 weakly excites
- L23:2491 weakly suppresses
- L13:437 weakly excites end of a short definition or claim
- L14:501 weakly excites
- L12:1725 weakly suppresses
- L14:285 weakly excites
- L14:548 weakly excites
- L12:959 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
- мед
- enn
- enne
- ญา
- 甑
- Convertible
- -aged
- Perl
- |
- Verb
Suppresses
- $⏎
- arten
- $ ⏎
- $⏎
- $")⏎
- swana
- bjerg
- @⏎
- uces
- __()⏎