pipe before a table field label
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. 12 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 vertical bar that starts a new row in a table of facts, such as the bar just before 'Number of pages' or 'Owner'. When it was turned up, the model's writing did not change in any clear way that matched 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.
-
portal/species/overview?p_p_spp=257336⏎This map is based on the
-
|The Physical Object|⏎|Number of Pages||136|⏎Learn why the Common Core is important for
-
Place of Publication||Perth, Western Australia|⏎|Number of pages||173|⏎|ISBN (Print)||9 7
-
flag staff|⏎|- elevation||15.75 ft (4.8 m) over stone base|⏎July 3, 1929⏎|Owner||National Park Service|⏎|Easiest access||walkway from avenue
-
you will need the Adobe Digital Editions app, or BlueFire Reader or Txtr app. These are free, too. For more details see this article.⏎|Size: ||359 KB|⏎|Date published: ||
-
an exception, which is life as usual for a Python programmer.⏎if not server_response['successful']:⏎etag = server_response['file_handle'].info().getheader('ETag')⏎xml_document = server_response['xml_document']<|endoftext|>Deepak V
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 thinking about... I have been thinking... I have been thinking... I have been thinking... I have been thinking... and I have been thinking... I have been thinking... and I have been thinking... I have been thinking...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
0 00:10:22 CEST 2010⏎On 9/9/10 4:39 PM, Baba wrote:⏎> In below code "
-
December 6, 2010 %9 text/html %U http://blogs.discovermagazine.com/badastronomy/2008/10/13/death-by-meteorite/⏎Disclaimer: ComPAD
-
' Classroom.|⏎|Levels:||Late Elem. (3-5), Middle School (6-8), High
-
and then its right child.⏎T / \ I S / \ D E<|endoftext|>Original URL: http://www.theregister.co.uk/2006/08/21/mars_geysers/⏎Martian pole freck
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
- L11:2834 suppresses
- L4:536 suppresses
- L4:19 weakly suppresses
- L6:398 weakly excites
- L8:2316 weakly suppresses
- L2:2791 weakly suppresses
- L11:2352 weakly suppresses
- L10:720 weakly suppresses
- L9:2433 weakly excites
- L2:1055 weakly excites
- L11:1174 weakly excites
- L1:2934 weakly excites
- L4:1040 weakly excites
- L11:253 weakly excites
- L8:193 weakly excites
- L7:1860 weakly suppresses
- L11:2089 weakly suppresses
- L11:515 weakly excites 'hands' in hand-washing advice
- L11:417 weakly excites
- L9:120 weakly excites
- L3:2135 weakly suppresses
- L5:1594 weakly excites
- L4:2992 weakly excites
- L7:315 weakly excites
Feeds later layers
- L16:1124 excites
- L13:2141 excites end-of-text token after finished articles
- L16:1913 excites
- L20:904 excites
- L19:2132 excites
- L19:1717 excites
- L13:901 excites
- L20:941 excites
- L14:1517 excites
- L15:2090 excites
- L16:2988 excites
- L16:1790 excites
- L21:738 weakly excites
- L19:870 weakly suppresses
- L17:882 weakly excites
- L13:625 weakly excites
- L18:1502 weakly excites
- L17:305 weakly excites
- L18:2204 weakly excites
- L17:2079 weakly suppresses
- L22:319 weakly excites
- L18:557 weakly excites
- L16:1447 weakly excites
- L15:1673 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
- ewidth
- anggal
- scope
- rieg
- 啥
- scope
- AÇÃO
- lement
- lijk
- göre
Suppresses
- 串
- 就够了
- 镑
- .getOutputStream
- öst
- othy
- compromises
- ((__
- ören
- ByteArrayOutputStream