line break inside a sentence
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 a line break that cuts a sentence in two, most often in scraped web text with awkward wrapping. Turning it up gave no clear theme in the output.
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.
-
MR file formats does not make for very exciting⏎reading. On the other hand many NMR spectroscopists have experienced the⏎frustration of having collected perfectly good spectra but not been able⏎to visualize
-
in three stages. The first stage⏎is to identify market trends and developments, including contact by⏎the designers directly with the market; some are stationed in toy⏎shops close to holiday periods, while others interview children. The⏎second stage is
-
Cattle truck and the Sheep⏎truck change places and the Engine is back on the main line?⏎In how many ways can you fit two of these yellow triangles⏎together? Can you predict the number of ways two blue triangles can⏎be
-
designed to have students analyze energy use, cost, and source patterns from household to regional scales and relate these patterns to CO2 emissions. This idea was generated at ...⏎Learn more about this review process.⏎Mentos and soda eruptions- lessons
-
total carnitine⏎blood levels, dicarboxylic aciduria.⏎Mild hyperammonemia and slight elevation of liver transaminases may also⏎be present with fatty acid oxidase disorders. Final diagnosis is established⏎by demonstration
-
an inorganic⏎substance, having an orderly atomic structure⏎and a definite chemical composition. The basic⏎source of all minerals is the hot magma in the⏎interior of the earth.⏎(more content follows the advertisement below) A D
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 have been following up with the user to make sure the user's actions to ensure that the user is clear and concise and accurate. to be clear and accurate and accurate and concise and clear and accurate and accurate and concise and accurate and clear
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
spread out or not concentrated in one place. Diffuse can also be a verb meaning to circulate or to spread.⏎- She is the CEO of a large diffuse company. (Here, diffuse is an adjective meaning spread out.)⏎- I need
-
⏎ones later to<|endoftext|>The “Where the Hell is Matt 2008?” video is a great example of the impact of music on images and changes how we look at it.⏎In case you haven’t seen this,
-
the Cham language was in full use. Most preserved manuscripts focus on religious rituals, epic battles and poems, and myths.⏎Modern Chamic languages have the Southeast Asian areal features of monosyllabicity, tonality, and glottal
-
. Using eight UK data sets (9–21 years), we investigate changes in restoration success for (1) arable reversion sites, were grassland was established on bare ground using seed mixtures, and (2) grassland enhancement
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
- L3:590 excites
- L5:381 suppresses
- L4:337 excites
- L4:1044 excites
- L3:414 excites
- L4:32 suppresses
- L4:156 excites
- L6:161 suppresses
- L2:583 weakly suppresses
- L6:2905 weakly suppresses
- L3:59 weakly excites
- L2:1139 weakly excites
- L4:39 weakly suppresses
- L3:79 weakly excites
- L3:439 weakly suppresses
- L5:128 weakly excites
- L2:1580 weakly suppresses
- L3:860 weakly suppresses
- L6:1070 weakly excites line break inside a sentence
- L3:42 weakly excites
- L6:1069 weakly suppresses
- L0:215 weakly excites
- L3:240 weakly excites
- L2:295 weakly excites
Feeds later layers
- L8:2421 strongly suppresses
- L8:297 suppresses
- L8:1543 suppresses
- L9:1539 suppresses
- L8:1664 suppresses
- L16:795 suppresses
- L8:1443 suppresses
- L10:258 suppresses
- L15:1448 suppresses
- L10:2315 suppresses line break inside wrapped text
- L9:1881 suppresses
- L9:1116 suppresses
- L15:1466 suppresses
- L17:1242 suppresses
- L15:985 suppresses line breaks inside wrapped text
- L9:234 excites line break inside hard-wrapped text
- L8:1821 suppresses
- L17:910 suppresses
- L16:317 suppresses
- L15:2148 suppresses
- L12:2649 suppresses
- L8:625 suppresses
- L13:2753 suppresses
- L18:1105 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
- tic
- ulet
- flare
- rias
- гин
- atta
- plain
- platz
- burger
- -alpha
Suppresses
- 我喜欢
- kicker
- //}⏎
- 却发现
- Lorem
- .SetString
- 那份
- \Helper
- 帮我
- 一直处于