line breaks 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. 3 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 the invisible line break inside a sentence, where a line of text has been wrapped or split in the middle of a thought. The examples mostly come from text copied from PDFs or web pages with hard line wraps. Turning it up did not produce a clear theme in the replies.
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.
-
Understanding antonyms is part of vocabulary development and this worksheet offers good practice with antonyms there are some synonyms and other words mixed in too so it s important that This video first appeared as a facebook live on 15th october download a
-
actual practice that difficulties could⏎be resolved by discussion as they could not be resolved by force. Moreover,⏎the British Commonwealth was founded on the conception that war between⏎
-
as the uprising⏎of the British Empire. But we have no right to this position of special⏎and honorable responsibility unless we are determined to use it to the⏎full, not for our own
-
has said⏎that he would not have one Christian missionary in India if he rose to⏎power. And if the Zionist Jews became the governors of Palestine all missionary⏎work there
-
4. While the crops⏎could produce incredible yields when properly managed and with plenty of⏎water, doing so reduced the water table, added salt deposits to the⏎land, and made portions of the land unsuitable for
-
policy shifts in member states,⏎such as Greece, Italy the United Kingdom and Ireland. The Multireader⏎project examined complex navigation through digital materials, and the⏎TEDUB project covered a similar area for graphical information
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: the importance of the teacher., and a student, ready to help me with. Now, I need help here. Because of the the hard fact, I need. So, I have to help with. This is a very
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
it was her own foolish choice and actions that lead to her and Romeo's death. Though the death of Romeo and Juliet was a tragic love story of two star-crossed lovers, caused by their families and friends, more specifically Friar Lawrence,
-
580.3 m (1,904 ft), bears the name of the Scot who discovered it in 1880. William Sutherland was one of several Scots who explored the country in the 19th century
-
for those that don’t have the ability to move to another state, the college they wish to attend is just a few mouse clicks away.⏎Is this where grade-school level education is headed? I can’t say for certain that is the case,
-
shapes - single, semi-double and double-flowered - is enormous. There are geraniums that can reach a height of one metre, but also creeping groundcover which only grows to a height of 10cm and hanging varieties which can
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
- L2:2324 suppresses
- L1:1279 suppresses
- L2:323 excites
- L2:835 suppresses
- L2:1580 suppresses
- L2:2665 suppresses
- L1:557 excites
- L0:37 excites
- L1:999 weakly excites
- L2:1933 weakly suppresses
- L2:605 weakly suppresses
- L2:204 weakly excites
- L2:2786 weakly excites
- L1:428 weakly suppresses
- L1:1723 weakly suppresses
- L2:962 weakly excites
- L2:772 weakly suppresses
- L0:602 weakly suppresses words right before a number
- L0:538 weakly suppresses
- L0:2996 weakly suppresses
- L1:1960 weakly excites
- L2:2553 weakly suppresses
- L1:198 weakly suppresses
- L2:1486 weakly excites
Feeds later layers
- L4:67 suppresses
- L7:1666 weakly suppresses
- L8:2086 weakly excites
- L4:11 weakly excites
- L21:2755 weakly suppresses
- L13:2826 weakly excites
- L19:180 weakly excites
- L19:2223 weakly excites
- L22:1162 weakly suppresses
- L18:3005 weakly suppresses
- L9:1207 weakly suppresses unclear: mid-sentence tokens in web text
- L17:2460 weakly excites
- L8:2520 weakly suppresses
- L11:1847 weakly suppresses
- L7:2140 weakly suppresses
- L6:2546 weakly excites
- L13:226 weakly suppresses
- L6:1487 weakly suppresses
- L6:1799 weakly suppresses
- L9:2479 weakly suppresses
- L6:429 weakly suppresses
- L23:2777 weakly suppresses
- L16:2175 weakly excites
- L11:1369 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
- unting
- 所
- nię
- 权
- _structure
- usher
- uture
- uktur
- 位
- SIZE
Suppresses
- 一听
- ypi
- 鸨
- ¥
- /GL
- Thief
- raid
- 杀手
- Decay
- reator