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. 18 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 lights up on the dash in words that are joined by a hyphen, such as 'high-yield' or 'RNA-Seq'. When it was turned up, the model's output turned into a jumble of word fragments with no clear meaning.
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.
-
wigeon and scaup, teal and shovelers, buffleheads and a few mottled ducks.’⏎- ‘The sea ducks - common golden-eyes and buffleheads - are up in central Alaska.’⏎- ‘The
-
counterpart–something that’s key to the preservation of ecosystems and their biodiversity.⏎By comparing the impact of wheat, rice, beef, and dairy production in both high- and low-yield farming systems globally, the authors found that in some cases high
-
15393 (393 Red).⏎Fedorets saw McConnell.⏎Captain Fedorets had just shot down F-86A-5-NA 49-1297, flown by 2nd Lieutenant
-
, zinc, flavonoids, B vitamins, and fiber. It is also low in calories.⏎Nutritional Value and Health Effects⏎A 100-gram (3.5 oz) serving of oca provides almost 40
-
to create a highly-competitive economy.⏎Even if we were to accept participation in the economy as the highest
-
outputs are high, resulting in a wired-AND function of the outputs. The product-term line is low when any of the
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 want to make a small but very important part of a project in the following way as a way to keep the overall project as a more __ ( fill the blank with the appropriate word in the sentence to complete the sentence in the sentence.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
9/10 grade band of the Minnesota Language Arts standards are addressed.⏎Credit: 1.0 (Semester A = 0.5 credit, Semester B = 0.5 credit.)⏎Honors Opportunity: Yes⏎Pr
-
1921 as Ireland was in the midst of the ‘War<|endoftext|>According to a recent study from the Journal of the American Medical Association (JAMA. 2010;303(18):184
-
10, with 96% of cases reported in three districts. There are numerous policy challenges to providing an effective public health response system in a fragile state<|endoftext|>Overview of atrial flutter⏎- Robert Phang, MD, FACC,
-
Sort: Correlations⏎Your teacher will give you a set of cards. Match each scatter plot with a pair of variables. Be prepared to explain your reasoning.<|endoftext|>The Saturday Evening Post turns 200 this year, so we are bringing
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
- L15:59 excites
- L16:170 suppresses
- L14:348 suppresses
- L9:7 excites
- L15:241 excites
- L14:51 excites colon that introduces a statement
- L17:620 excites
- L14:422 excites periods after initials and abbreviations
- L17:2829 excites
- L15:286 excites
- L17:382 suppresses closing bracket after a fact in parentheses
- L10:33 suppresses
- L16:337 weakly excites
- L15:152 weakly suppresses
- L7:92 weakly suppresses
- L16:258 weakly excites 'of' after 'all' in phrases
- L12:2736 weakly suppresses
- L5:5 weakly suppresses digit after a label like paper or type
- L5:238 weakly suppresses
- L13:1046 weakly suppresses 'in' or 'on' ending a phrase
- L16:955 weakly excites
- L17:773 weakly suppresses
- L16:314 weakly excites
- L6:18 weakly excites
Feeds later layers
- L20:59 strongly excites
- L26:21 strongly suppresses
- L22:14 strongly excites
- L22:16 strongly suppresses comma in dates and birth lines
- L25:30 strongly excites
- L20:489 strongly excites
- L25:17 strongly excites
- L24:51 excites
- L21:70 excites
- L25:109 suppresses
- L23:80 excites blank slot before a year or number
- L27:33 excites
- L20:193 suppresses
- L26:48 excites
- L27:11 suppresses unclear: mixed fragments in lists and pronunciations
- L19:509 excites
- L21:214 excites
- L27:97 excites
- L21:118 excites
- L27:85 excites
- L24:39 excites
- L21:418 excites
- L21:29 excites
- L21:230 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
- -sama
- وبة
- ísticas
- kHz
- =params
- 瓠
- %;"
- usan
- 焦点
- 桌
Suppresses
- e
- me
- rate
- One
- Re
- re
- one
- ?
- None
- even