hyphen joining 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. 19 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 a hyphen that links words into one compound, like 'mini-satellites', '16-lead' or 'tick-removing'. Turning it down made the model spit out long runs of hyphens, so this neuron is clearly tied to the hyphen mark.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
, about 19 percent of 25- to 34-year-olds had a college degree, compared with 35 percent of 35- to 44-year-olds and 38 percent of 45-
-
the U.S. and abroad. The Federal Emergency Management Agency (FEMA), continuously works with local governments, states, and tribes regarding hazard mitigation plans with short- and long-term focus. These efforts have the overall intent of increasing education and awareness
-
designed to support orbits for the satellites, to perform orbit control tasks for small-, mini- and micro-satellites.<|endoftext|>Until 1864, Georgia had
-
60 second-, third-, and fourth-grade students, while 48 second- and fourth-grade students were subjects in the other 3 studies. Results indicated that
-
.It is in the latter provokes respiratory failure and signs of pulmonary heart.⏎should be noted that asthma in heavy current, emphysema, and other long-occurring diseases are the causes that resulted in developing chronic obstructive pulmonary disease.⏎
-
2. The ADP1613 is available in an 8-lead MSOP package and the AD5504 in a 16-lead TSSOP.⏎The boost circuit can also be used in LED driver applications
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.
-
effect on their decision. Today, the widespread use of the Internet and smartphones has enabled consumers to collect extensive information via viral marketing. This study aimed to compare word-of-mouth (WOM) and viral marketing strategies on consumption patterns, information traits,
-
be interpreted as a one-word question: “Tasty?”⏎And here is the rest of the story, a special trail tale that takes visitors to a one-of-a-kind place.⏎I am amazed at
-
and Bryan Cervical Disc are approved for distribution in the United States by the FDA.⏎An Alternative to Traditional Spinal Fusion⏎Age, genetics and everyday wear-and-tear of routine activities eventually can contribute to damage and degeneration of
-
as well as the presence of a virus, disease, or food-related allergy, should be ruled out. A step-by-step approach should be used to identify a possible DISR:⏎1. Complete a
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
- L18:691 strongly suppresses
- L16:844 suppresses
- L4:46 suppresses
- L17:620 excites
- L16:363 excites
- L17:364 excites hyphen joining two words
- L1:320 excites
- L15:241 excites
- L18:375 excites
- L0:1703 excites
- L15:1042 excites
- L0:848 excites
- L4:39 suppresses
- L14:136 excites
- L14:732 weakly excites
- L8:475 weakly suppresses
- L17:603 weakly excites inside organisation and product names
- L10:156 weakly suppresses
- L16:406 weakly suppresses
- L4:103 weakly excites comma after 'still' or 'in addition'
- L3:101 weakly suppresses
- L16:627 weakly excites
- L4:1300 weakly suppresses
- L10:830 weakly suppresses
Feeds later layers
- L22:494 strongly suppresses
- L25:149 strongly suppresses
- L26:149 strongly suppresses
- L27:226 strongly suppresses
- L27:138 strongly suppresses
- L20:1571 strongly suppresses
- L26:596 strongly suppresses
- L27:1459 strongly suppresses
- L23:1652 strongly suppresses
- L25:1030 suppresses
- L24:266 suppresses
- L26:811 suppresses hyphen in number or size compounds
- L21:321 suppresses
- L21:378 suppresses
- L26:2789 suppresses
- L20:317 suppresses
- L26:664 suppresses
- L22:189 suppresses
- L25:31 excites
- L20:1064 suppresses
- L24:454 suppresses
- L23:793 suppresses
- L27:1098 suppresses
- L24:326 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
- however
- 但是
- however
- 但是对于
- 网讯
- 则是
- ting
- though
- tant
- 却是
Suppresses
- -
- -that
- -pro
- -W
- -\
- -cur
- -j
- -v
- -",
- -a