commas between numbers in lists
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. 27 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 a comma that sits between numbers in a list or sequence, such as '1,2,3,4' or '143,754', and to a lesser extent on commas between items in a list. Its negative side is weak and comes from ordinary commas in prose. Turning it up made the text repeat words like 'much', and turning it down produced repeated Chinese characters, so the steering results do not clearly reflect the pattern.
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.
-
still useful. Arabic numerals share their origin from the numbering system of India.⏎The pattern of Arabic numerals start from 0,1,2,3,4 and it simply goes up in such way. On the other hand Roman numer
-
icata|⏎|Species Authority:||Todd, 1913|⏎|Red List Category & Criteria:||Endangered B1ab(i,ii,iii,v) ver 3.1|⏎|Reviewer(s):||Butchart
-
,182; Aurora, 286,028; Lakewood, 143,754; Fort Collins, 124,665;
-
link="none" columns="2" size="medium" ids="2568,2567,2569,2570"]These
-
couch or the bed. Shelter,food,clothes,education, and family are all the things that every person deserves. These are all basic needs
-
Previous: [SVG] How to draw a rectangle?⏎<svg height="210" width="500"> <polygon points="200,10 250,190 160,2
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:并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
about how to correct material in RePEc.⏎For technical questions regarding this item, or to correct its authors, title, abstract, bibliographic or download information, contact: (Michael McNulty)⏎If references are entirely missing, you can add
-
提要 [ summary or outline of a book, play, etc ]⏎note: 字根組合的概念是 "general view",可以用來指電視劇、書<|endoftext|>Making the decision for Google to change google.ps’ website to mention “
-
法、分析法来进行研究,探讨泰国学生在学习汉语助动词中的困难及偏误情况,分析他们产生偏误的原因。我们发现经过基于对比研究设计的汉语
-
Learners Based on the Comparative Analysis⏎- 人文学院-已发表论文 ⏎“助动词”在汉语中具有较大的作用,使用率高,但是泰国学生经常在学习和使用汉语助动词中产生各种的
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
- L25:535 strongly excites commas inside number lists
- L25:1757 strongly suppresses
- L26:1876 excites
- L20:1298 suppresses
- L20:1385 suppresses
- L15:184 excites
- L19:43 suppresses
- L19:1826 suppresses
- L25:422 excites
- L18:251 excites
- L26:75 suppresses
- L10:155 excites
- L18:66 suppresses
- L13:44 suppresses inside web addresses and citations
- L24:248 excites
- L4:3 excites commas after citation fields
- L16:7 excites
- L2:422 suppresses
- L16:240 excites comma after 'sometimes' or 'the truth'
- L5:97 excites
- L14:35 suppresses
- L25:231 weakly excites
- L0:132 weakly excites
- L18:2403 weakly excites
Feeds later layers
None: this is the last layer.
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
- inals
- .ravel
- Pieces
- (IN
- ائية
- ians
- .flatten
- agger
- eldorf
- much
Suppresses
- 并
- 则
- 并且
- 并将
- 请
- 并对
- 在
- 并通过
- 请你
- 并于