unclear: mid-word splits in foreign or code text
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. 8 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 odd word pieces, like the middle of a split word or a fragment inside a web address or a foreign word. When it was turned up, the model's writing drifted toward topics like school and technical help, and it got less coherent.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
years later<|endoftext|>Hand-coloured lithograph entitled; 'Dampfmaschinen Pferd worauf man in einer stunde von Paris nach Petersburg reiten kann' ('Steam engine horse on which one can
-
ulunda tamamlamıştır.⏎Yüksek Lisans Öğrenimine MSÜ Barbaros DEBİM Enstitüsünde devam etmekte aynı zamanda enerji sistem ve modelleme üzerine araştırm
-
those involved in educational leadership and policy development. This work is also valuable for those interested in the locally organized and interactionally achieved context of institutional work' - Discourse:
-
7075. . . http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0057075.<|endoftext|>Racism
-
Sustainability 2013, 5(7), 3006-3008; doi:10.3390/su5073006⏎I am honored to introduce this
-
yellow. The french blog Volcanic News Erutpion states it as being restless. http://volcansvanuatueruptionsgb.blogspot.fr/2014/03/08032014
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 need to make a 200/300/300/300 configuration for a home-made 200/300/300 configuration. The first part is 200
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.⏎For intramuscular use only.⏎Recommended Dose And Schedule⏎The dose and schedule for Fluvirin is presented in Table 1.⏎TABLE 1 : Fluvirin Dose
-
enclosure is designed to be corrosion and UV resistant and strong enough to withstand falling trees. The electronics within the enclosure are shown in Figure 6.⏎Figure 5. Typical element enclosure.⏎Wind speed, wind direction, and temperature are also recorded at
-
The severity of fractures with large pieces of bone dislodged indicate that these were formed from violent blows in a standardised manner (Figure 1).⏎Figure 1: Top row, Cranium 20
-
, but every year the Internet packet transmission market could grow as shown in Figure 4-1.⏎Figure 4-1. The unit cost of Internet bandwidth drops, but the volume increases.⏎The result
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
- L6:1601 excites
- L7:2701 weakly suppresses
- L2:67 weakly excites
- L3:23 weakly suppresses
- L5:1938 weakly suppresses
- L5:1319 weakly suppresses
- L2:2771 weakly excites
- L7:1972 weakly excites
- L1:168 weakly suppresses
- L2:2202 weakly excites
- L6:1102 weakly excites
- L2:1465 weakly excites
- L5:919 weakly suppresses
- L0:2304 weakly excites
- L4:1963 weakly excites
- L0:2769 weakly suppresses
- L7:3029 weakly suppresses
- L6:2364 weakly excites
- L4:245 weakly excites
- L6:2479 weakly excites
- L7:1970 weakly suppresses
- L3:2884 weakly excites
- L1:1272 weakly suppresses
- L2:2956 weakly suppresses
Feeds later layers
- L10:1874 suppresses
- L10:586 excites 'important' and degree words after 'how'
- L9:799 suppresses
- L19:1361 weakly excites
- L14:1498 weakly excites
- L20:810 weakly suppresses
- L9:288 weakly excites
- L13:1513 weakly excites
- L16:1173 weakly excites
- L25:1693 weakly suppresses
- L15:1955 weakly suppresses
- L16:1036 weakly excites
- L19:2756 weakly suppresses
- L23:1186 weakly excites
- L19:2575 weakly suppresses
- L14:827 weakly suppresses open bracket after a value or citation
- L18:104 weakly excites
- L21:1035 weakly suppresses
- L23:2273 weakly suppresses
- L19:45 weakly excites
- L13:1001 weakly suppresses
- L27:1244 weakly suppresses
- L11:687 weakly excites
- L20:2816 weakly 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
- anes
- 萦
- uada
- apel
- เอง
- 之势
- 羌
- ושר
- 都要
- ower
Suppresses
- :///
- "<?
- 读后
- reads
- boa
- ented
- ento
- ello
- ęp
- adays