dash in year or number ranges
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. 7 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
Fires on the dash that joins two years or numbers, like 2011-2012 or 15-19. Turning it up did not produce any clear theme in the writing.
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.
-
with the Apiary Inspectors of America and The Bee Informed Partnership.⏎The losses were 42% higher than during the 2011-2012 winter, when 21.9% of colonies died
-
-- and an early one. While the article admits that scientists and doctors struggle to predict flu seasons each year, general consensus indicates that the 2017 - 2
-
academic needs, as well as any social or emotional concerns they’ve noticed.⏎The high school has been implementing the BARR program since the 2017-2018 school year and has seen significant improvements in course credits earned,
-
2017-18 Extension Wheat Production Guide, Georgia farmers planted 160,000 acres in the 2016-2017 season, which marked the third year the state’s wheat acre
-
DOE, the Department of Defense and NASA. The company hopes to raise another $2 million by next year, $35 million in 2015-17, and $200 million for its pilot plant stage.⏎The
-
, that parents should personally educate their kids. The Scheme of Inclusive Education for Disabled at Secondary Stage (IEDSS) was launched throughout 2009-10 and replaces the earlier scheme of Integrated Education for Disabled Kids (IEDC
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: this is the time of year, and I have to do this. It's important. This is a problem. So. Let's go. It's a good day. Let's not do it. It's the right choice. The
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
vomiting (p < 0.001, OR = 1.969, 95% CI = 1.618-2.398), while fewer manifested fever (p = 0.
-
ks (p = 0.001, OR = 1.348, 95% CI = 1.124-1.618) were affected with norovirus when compared with non
-
.73–0.97, n = 25) and 0.72 (95% CI = 0.52–0.86; n = 27), respectively. Results indicated that
-
(95% CI, 1.4-3.0) for alcohol, 4.2 (95% CI, 3.1-5.8) for marijuana, and
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 suppresses
- L3:101 suppresses
- L6:1868 excites
- L5:291 excites hyphens inside codes and page ranges
- L1:404 weakly suppresses
- L1:320 weakly excites
- L3:135 weakly suppresses
- L6:831 weakly excites
- L4:799 weakly suppresses
- L2:18 weakly suppresses
- L6:2170 weakly excites
- L4:25 weakly excites
- L6:3009 weakly excites
- L2:66 weakly suppresses
- L5:1817 weakly suppresses
- L3:807 weakly suppresses comma inside dates and years
- L0:1911 weakly suppresses
- L1:2234 weakly suppresses
- L5:2251 weakly suppresses
- L6:1235 weakly suppresses
- L6:900 weakly suppresses
- L6:3050 weakly suppresses
- L4:1170 weakly suppresses
- L6:2159 weakly suppresses
Feeds later layers
- L12:1851 excites hyphen inside dates
- L9:1820 excites
- L17:364 excites hyphen joining two words
- L14:2202 excites
- L19:804 excites
- L9:589 excites
- L15:2247 weakly excites
- L15:1042 weakly excites
- L15:1549 weakly excites
- L9:1031 weakly suppresses
- L22:1331 weakly excites
- L19:2462 weakly excites
- L16:2498 weakly excites
- L16:2580 weakly excites 'and' linking food or health nouns
- L9:392 weakly excites "and" inside organisation names
- L11:2394 weakly excites
- L24:201 weakly suppresses
- L19:1475 weakly excites
- L10:1367 weakly excites
- L20:1886 weakly excites
- L8:1913 weakly excites
- L21:1624 weakly excites
- L16:1246 weakly excites
- L17:2613 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
- մ
- み
- veys
- 什么呢
- どう
- änn
- icators
- experience
- �
- 你看
Suppresses
- enville
- ственно
- 后果
- disadv
- idar
- 独角
- dra
- =&
- .rawValue
- setResult