end-of-text token after a document
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. 0 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 marker the training data uses to separate one web page from the next, which shows up right where a document ends. Its strongest negative reactions are on a closing quotation mark followed by a comma. Steering had no clear visible effect.
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.
-
their . place. So, the students, teachers and other professionals finally had to resume their studies and jobs in government institutions like schools, colleges, courts etc.<|endoftext|>I hope you are all well and that you enjoyed some of your weekend in the
-
Hydrogen<|endoftext|>Educational Technology and Mobile Learning: 8 Excellent Augmented Reality Apps for iPad
-
management, risk management and time management skills.⏎- Software development in Java using various web-development technologies.⏎Figure 1: Traceability User Interface⏎Back to top<|endoftext|>William Gilly has seen a Kraken. The mythical squid beast with⏎ship
-
which can be physical or concepts, etc⏎Definitions of Rail:⏎Cambridge Dictionaries |⏎New Zealand |⏎Similar Words (using 'rail' as word stem):<|endoftext|>First Flag Flown⏎In 1960, when CDC moved from
-
Saturday<|endoftext|>15.1: The Science Fair (5 minutes)⏎The purpose of this
-
Rocky Mountains. OSC will continue to better understand wolverine behavior and advocate for increased local and national coordination.<|endoftext|>Regional Broadband Global Area Network (RBGAN)⏎Definition - What does Regional Broad
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 trying to understand the idea of a system, and I have been thinking about it. I have been thinking about it, and I have been trying, and I have been thinking about it. I have been thinking about it,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
,” according to Currier.⏎Read the full article, Which States Have the Most Economic
-
,” Dawson said. “Alltech’s distinct set of technologies can help in maintaining healthy
-
,” we get the subject form (An Luan), and, by following genitive
-
,” 05/31/2002, WGBH Media
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
None: this is the first layer.
Feeds later layers
- L2:39 excites
- L2:54 excites
- L1:1929 weakly suppresses
- L1:110 weakly suppresses
- L1:2677 weakly suppresses
- L1:1321 weakly suppresses
- L2:1489 weakly suppresses
- L1:2480 weakly excites
- L25:478 weakly excites
- L2:46 weakly suppresses
- L10:371 weakly suppresses
- L2:60 weakly suppresses
- L2:212 weakly suppresses
- L9:623 weakly excites
- L2:128 weakly suppresses
- L1:37 weakly suppresses
- L4:239 weakly suppresses
- L2:55 weakly suppresses
- L1:2159 weakly suppresses
- L1:1262 weakly suppresses
- L1:868 weakly suppresses
- L14:1123 weakly suppresses
- L19:755 weakly suppresses
- L11:801 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
- 小
- présent
- ,label
- gis
- ,[
- 一个小
- Sm
- grim
- ogr
- uler
Suppresses
- impse
- 但这
- Paramount
- ))?
- NESS
- yet
- ariate
- surf
- обще
- 阗