document break token mid-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. 21 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 reacts most strongly to the invisible marker that separates two unrelated documents, especially when it lands in the middle of a sentence. Its negative side reacts to 'i.e' abbreviations. Turning it up or down mostly made the text degrade into repeated punctuation or unrelated talk, so the steering shows no clear theme.
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.
-
of workers to a consultancy. These consultancies offer comprehensive services, which normally include –⏎• Activities related to registration and licensing necessary under the labour laws,⏎•<|endoftext|>- Making Maths Accessible⏎Version Dated 24th February 2
-
rounded to the nearest integer.<|endoftext|>Christine Duncan, MD⏎Sickle cell disease (SCD) affects as
-
a phase boundary, for instance, gas-oil, oil-water, and<|endoftext|>Green kitchen: A recipe for being more environmentally friendly⏎Cooks can do their
-
they eat, so to avoid causing spikes in their blood glucose, but type 1 cannot be controlled solely with diet. The necessity for treatment with insulin is why<|endoftext|>Last updated: July 27, 2007July 2
-
chool.org/), our experiment with graphics quickly turned into a full-grown graphic anthropology workshop, and some of the participants used sketching in ways that did not<|endoftext|>People vary in their reactions to mosquito bites. Most people develop itchy, raised
-
. It is because the re-exported goods might require some additional information.⏎Having understood so much about reexport trade and what are re-exportation, it is<|endoftext|>10.2.2 Restarting TCP/IP⏎TCP/IP is a peer
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.
-
is a two-letter rating that tells you the type of engine the oil is designed for (i.e. gasoline or diesel) and the quality level. The “S” designation (
-
type 1 and also non insulin dependent diabetes or diabetes type 2.⏎Diabetes is worldwide in distribution and the incidence of both types of primary diabetes, i.e insulin dependent diabetes mellitus (IDDM) or diabetes mellitus type 1
-
. For example, if the array is [1, 2, 3, 4, 5], it will print it in reverse order i.e. from the last element to the first element:⏎5 4 3
-
of selective toxicity: that certain chemicals that would be toxic to some organisms, e.g., infectious bacteria, would be harmless to other organisms, e.g., humans. In 1928, Sir Alexander Fleming, a Scottish
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
- L17:152 strongly excites mid-word fragments, often 'intric'
- L17:70 excites
- L6:293 suppresses
- L16:1421 suppresses
- L19:696 excites
- L10:43 excites "terms" in "in terms of"
- L12:50 suppresses
- L4:1572 excites
- L11:42 suppresses
- L7:9 suppresses
- L18:91 excites
- L14:102 excites
- L19:220 suppresses question mark ending a web question heading
- L5:2525 suppresses
- L8:62 suppresses
- L19:1708 excites
- L19:377 suppresses
- L20:131 suppresses
- L14:56 suppresses document ends and list line breaks
- L1:78 excites
- L5:34 excites
- L20:472 excites
- L6:1031 excites
- L5:1127 excites
Feeds later layers
- L27:156 suppresses
- L24:94 suppresses
- L25:30 suppresses
- L24:385 suppresses
- L24:39 suppresses
- L26:356 suppresses
- L27:151 suppresses
- L26:21 excites
- L22:765 suppresses
- L23:191 suppresses
- L22:14 suppresses
- L23:676 suppresses the 'm' in a.m. and p.m
- L22:424 suppresses
- L22:65 suppresses
- L24:34 excites
- L24:31 excites
- L24:508 suppresses
- L25:17 suppresses
- L23:80 suppresses blank slot before a year or number
- L27:11 excites unclear: mixed fragments in lists and pronunciations
- L27:301 suppresses
- L27:342 suppresses
- L26:131 excites
- L22:1733 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
- ,
- �
- not
- `
- **
- \
- se
- IA
- **
- and
Suppresses
- @update
- uai
- ableObject
- 子里
- ',...⏎
- etting
- ']->
- usses
- 了他的
- ",-