hyphen inside compound 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. 20 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 responds to the hyphen in words joined by a dash, like "above-mentioned," "at-your-own-pace" and "low-and middle-income." It is mostly silent elsewhere. Turning it up did not produce any obvious theme in the text that followed.
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.
-
COVID-19 diagnostic based on micro/nano technology and how to tackle the above-mentioned challenges.⏎The exploration of surfaces, interfaces, nanostructures and thin films using
-
honor the format of exploring hands-on activities, learning about artwork in our galleries, and making an inspired art project.⏎As always, Kid Quest is an at-your-own-pace art adventure. In this case, take all the time you like
-
The currently known individual allergens of the above-mentioned allergy triggers from the arthropod realm are summarized and their potential use in
-
engineers will hold their breath as the newest Landsat satellite is launched into space aboard an Atlas V rocket. Landsat 8 carries one of the most sophisticated Earth-observing systems, but it also carries the weight of a
-
Not at all, says Wolf. Instead, she argues, “The great challenge now is to learn how to use both print and digital mediums to their best<|endoftext|>- A Tale of Two Brothers: Haiti’s Other Revolutions⏎Sévère Cour
-
of the online version is:⏎Several examples of recently-constructed buildings are shown, all of them designed with the latest 3-D modeling
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: theseseseseseseseseseseseseseseseseseses.This is the first thing I have done, andI have the first callout callout callout
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Proceedings of the Iowa Academy of Science, 31(1), 131-132.⏎Available at: https://scholarworks.uni.edu/pias/vol31/iss1/25<|endoftext|>New images
-
.⏎Avesoro CRUSHING AND PROCESSING OPERATIONS AT NEW⏎2 Nov 2015 Ore crushing operations recommenced on 28 October 2015, following the “
-
Build a broad network for yourself. Seek to enlarge you network of associates, with the goal of expanding your interests and opportunities.⏎COMMUNICATE. NEGOTIATE. APPRECIATE.<|endoftext|>West Nile Virus⏎News of an epidemic
-
call came to late. Fedorets fired at McConnell with his MiG-15’s three autocannon (a single Nudelman N-37 37 mm gun and two Nudelman-Rikhter NR-2
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
- L2:404 strongly excites
- L19:89 strongly excites
- L1:217 strongly suppresses
- L3:508 strongly suppresses the final zero of decades like '1980s'
- L19:43 strongly excites
- L16:3 strongly suppresses
- L5:97 strongly suppresses
- L5:621 excites
- L16:7 suppresses
- L16:0 suppresses
- L3:395 suppresses
- L4:490 suppresses decade digit in years like 1930s
- L16:1 excites
- L15:123 suppresses
- L0:132 suppresses
- L8:3004 excites
- L0:2215 suppresses
- L2:801 suppresses
- L6:715 excites
- L2:761 suppresses
- L2:422 excites
- L18:853 suppresses
- L0:579 suppresses
- L4:384 suppresses
Feeds later layers
- L22:265 strongly excites
- L24:248 strongly excites
- L25:422 strongly excites
- L24:318 strongly suppresses
- L25:231 strongly suppresses
- L26:1008 strongly suppresses
- L27:2408 strongly suppresses
- L26:2003 strongly excites hyphen joining two word parts
- L26:254 strongly suppresses
- L27:185 strongly excites
- L27:1463 strongly excites
- L22:2384 strongly excites
- L26:1776 strongly excites
- L27:2439 strongly excites slash inside web addresses
- L27:145 strongly excites dots inside web addresses
- L27:1019 strongly suppresses
- L26:725 excites
- L27:170 excites unclear: spaces in mixed-language text
- L27:981 excites
- L27:2657 excites
- L23:474 excites
- L23:393 suppresses
- L25:36 excites
- L25:535 excites commas inside number lists
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
- RIPTION
- /he
- цин
- ẩn
- ائن
- ForResult
- Є
- Offsets
- 见过
- арь
Suppresses
- s
- y
- i
- t
- n
- a
- m
- S
- d
- sing