hyphen joining two words
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. 17 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 hyphen that joins two words into one compound, such as "cold-weather" or "tobacco-cessation." Its weaker opposite side fires on minus signs in numbers. Turning it up or down gave text that was not clearly about hyphens.
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.
-
know it (more or less). Thus began the era of dancing babies and other horrors best forgotten<|endoftext|>PROTEOMICS is the study of many proteins at once- or a description of proteins in a cell, tissue or cellular organelle on
-
. Hernias can be painful and limit the mobility of pigs.⏎- Having<|endoftext|>This time of year, it may be difficult to detect the cause of your cold-weather sniffles. Colds, sinus infections and the flu all share many symptoms
-
MIT News Office. 1992-04-15. http://web.mit.edu/newsoffice/1992/citation-0415.html. Retrieved 200<|endoftext|>Our herd of awesome
-
the participants.⏎People generally didn’t mind being touched by Cody overall, but were less comfortable with the robot when it spoke to them beforehand. And participants were more-accepting of a potentially necessary medical touch than an attempt at a soothing touch by
-
48 percent in Hispanics); and less likely to have used tobacco-cessation aids during the past year in a quit attempt (38 percent in
-
that, it is the way women-issues are treated at the working place that translates into them earning less than their male
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: - - _ 1. 1000 1000 1000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
i Marincas | Shutterstock⏎Atomic Number: 86⏎Atomic Symbol: Rn⏎Atomic Weight: 222⏎Melting Point: -95 F (-71 C)⏎Boiling Point: -79 F
-
This example demonstrates different noise distributions methods that can be selected for a Noise block. Both noise blocks use samplePeriod = 0.02 s, y_min=-1, y_max=3, and have identical fixedLocalSeed. This means
-
In July, the warmest month, the average summit temperature is -19° C (-2° F).⏎When George Mallory, the British
-
40°F (-30°C/40°C) Other cover qualities (such as oil or fire) are usually only able to withstand a minimum temperature of -20°C. For temperatures lower than this, conveyors should be fitted with
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
- L4:46 strongly suppresses
- L16:844 strongly suppresses
- L0:1703 strongly excites
- L3:101 suppresses
- L0:848 excites
- L16:893 suppresses
- L1:320 excites
- L15:241 excites
- L16:264 excites opening parenthesis in citations and asides
- L7:89 excites dash in year or number ranges
- L6:338 suppresses
- L1:38 suppresses
- L7:432 excites
- L5:1304 excites
- L1:1336 excites
- L0:1267 suppresses
- L10:823 suppresses
- L0:1128 suppresses
- L5:291 weakly excites hyphens inside codes and page ranges
- L10:156 weakly suppresses
- L14:415 weakly suppresses
- L2:317 weakly suppresses
- L10:134 weakly excites
- L7:122 weakly excites
Feeds later layers
- L18:691 strongly excites
- L19:804 strongly excites
- L20:1886 strongly excites
- L19:1879 suppresses
- L19:2267 excites hyphen joining compound words
- L18:1142 suppresses
- L18:642 suppresses
- L26:2003 excites hyphen joining two word parts
- L23:1479 excites
- L18:1541 suppresses
- L22:2211 excites
- L23:793 weakly suppresses
- L18:2369 weakly suppresses
- L19:1398 weakly suppresses
- L18:1047 weakly excites
- L19:171 weakly suppresses
- L24:266 weakly suppresses
- L25:2347 weakly excites
- L18:713 weakly suppresses
- L27:504 weakly suppresses
- L20:183 weakly suppresses
- L25:821 weakly excites
- L20:1571 weakly excites
- L26:811 weakly excites hyphen in number or size compounds
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
- -Identifier
- ual
- NOWLED
- çe
- ysis
- ance
- rench
- èle
- nowled
- addr
Suppresses
- раг
- عيد
- шел
- Form
- Forms
- den
- +"]
- bur
- 发
- 齊