end-of-document token after a web snippet
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. 11 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
Fires weakly on the invisible marker that the model uses to show one web page or document has ended and another begins. Turning it up made the model produce hesitant, rambling "I mean" text.
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.
-
dark brown (2).⏎- Length: 11 cm (2)⏎- 6 g (2)<|endoftext|>Causes of Cancer⏎Cancers can develop from a variety of sources and causes
-
per minute|⏎|45 years||88-149 beats per minute||175 beats per minute|⏎|50 years||85<|endoftext|>A Brief Introduction
-
: ||9781312744967 (DRM-EPUB)|⏎|Read Aloud: ||not allowed|<|endoftext|>In 1979, police in Pensacola, Florida, arrested
-
.org/catalog/V_4E4EE3EAD24043E2BC0B972CD24B176C<|endoftext|>Rurik DynastyArticle Free Pass⏎Rurik Dynasty, princes of K
-
2021⏎- Nepali Climbers Open New Trekking Route on Mt. Kilimanjaro - September 9, 2021<|endoftext|>In this VB.NET sample project, you will learn to create a simple dictionary.
-
In The Moon"⏎- 'Star Trek: The Next Generation' - "Rascals"⏎- 'Warehouse 13' - "Time Will Tell"<|endoftext|>I question why "pres" was marked incorrect in this exercise. As I understand
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: **************************************************
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
J.M.; Mortimer, S.R.; Brereton, T.; Redhead, J.W.; Thomas, J.A.; Pywell, R.F.. 20
-
different. Although the latest legs have exceeded all expectations in stress-testing–the 3-D-printed plastic is about as strong as a traditional prosthetic-–coons and associate professor Matt Ratto have yet to find the perfect thickness, the
-
2.16.128.0/17 SUBNET3 : 10101100.00010001.0yyyyyyy.yyyyyyyy => 172.1
-
/Share-Alike License⏎- n. An underwater explorer.⏎from WordNet 3.0 Copyright 2006 by Princeton University. All rights reserved.⏎- n. an underwater swimmer equipped with a face mask and foot fins
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
- L9:15 strongly excites
- L5:25 strongly excites
- L10:28 strongly suppresses unclear: weak, articles before nouns
- L5:36 suppresses
- L10:51 excites
- L5:112 excites
- L5:498 suppresses
- L10:150 suppresses
- L7:53 weakly excites
- L2:69 weakly excites
- L0:18 weakly suppresses
- L6:7 weakly suppresses
- L2:75 weakly suppresses
- L6:13 weakly excites
- L8:10 weakly excites
- L3:12 weakly excites 'and' joining the last item in a list
- L3:2870 weakly excites
- L5:110 weakly suppresses
- L5:46 weakly excites
- L6:24 weakly excites
- L3:115 weakly suppresses
- L8:14 weakly excites
- L6:18 weakly excites
- L10:895 weakly suppresses
Feeds later layers
- L13:654 strongly suppresses
- L13:105 strongly excites
- L14:847 strongly suppresses
- L14:68 strongly excites
- L14:210 strongly suppresses
- L15:892 strongly suppresses
- L13:2270 strongly excites
- L15:62 strongly excites
- L12:1574 strongly excites
- L12:396 strongly excites
- L15:514 strongly excites
- L15:907 strongly suppresses the '0' in numbers like 30 or 40
- L16:368 strongly suppresses
- L16:192 strongly suppresses
- L12:2159 strongly suppresses
- L12:1307 suppresses
- L18:942 suppresses
- L16:841 excites
- L12:741 suppresses
- L14:98 excites
- L14:2781 excites
- L15:1550 suppresses
- L15:13 suppresses
- L22:213 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
- отор
- ford
- ods
- dr
- dr
- utex
- Constructed
- [@
- utan
- onna
Suppresses
- grou
- bä
- ält
- 路线
- ACHER
- Pin
- â
- →person
- péri