chinese characters inside mixed 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. 19 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 responds to Chinese characters, especially inside text that mixes Chinese with English or other languages, such as a bracketed Chinese term or a blank in a classic poem. It is most active on the character itself. Turning it up gave English text with no clear Chinese link, so the steering did not reveal much.
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.
-
systems, it also means that pollutants that originate indoors are getting trapped inside.⏎“People normally<|endoftext|>Spring Dawn by Meng Haoran 春眠不觉晓 处处闻啼茑 夜来风雨声 花落知多少⏎
-
scholar, politician and philosopher who lived in the Spring and Autumn period. The main body of the Confucian canon comprises the Four Books and the Five Classics (四書五經), texts which have been traditionally attributed to Confucius himself,
-
be an accessible topic for middle school students.19<|endoftext|>毋忘在莒(Pinyin: Wu Wang Zai Ju) is a famous Chinese idiom (成語, Pinyin: Cheng Yu). It is sort of like the Latin id
-
udoki 肥前國風土記 (first part of the 8th century) 10. The reading (furigana 振り假名) of those texts, although, probably does
-
: place 闻 : hear 啼 : cry 茑 : bird⏎Meaning for Line 3 夜 : night 来 : come 风 :
-
the learner<|endoftext|>n.交響樂 [an elaborate instrumental composition in three or more movements, similar in form to a sonata but written for an orchestra (管弦樂隊) and usually of far grander proportions and more varied elements.
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 use of 360度点子云图图谱图谱子集集子集子集子集子集子集子集子集子集子集子集子集子集子集子集子
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
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。
-
to be an accessible topic for middle school students.19<|endoftext|>毋忘在莒(Pinyin: Wu Wang Zai Ju) is a famous Chinese
-
wiki: In the past, syndromes were often named after the physician or scientist who identified and described the condition in an initial publication⏎ex. 唐氏
-
being explored and the most valuable part of the ship (the stern castle, where the captain’s cabin would have concealed the most precious cargos of all) has yet to be located.⏎Priceless Knowledge: Devon’s Bronze Age wreck⏎Salcombe
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
- L18:811 excites
- L16:1293 suppresses
- L17:1590 suppresses
- L16:38 suppresses
- L17:257 suppresses
- L18:501 excites
- L9:146 excites
- L18:125 suppresses
- L8:763 excites
- L16:286 suppresses
- L18:1169 suppresses
- L8:423 suppresses
- L16:1432 excites
- L17:259 excites
- L3:305 excites
- L18:569 excites
- L11:1872 suppresses
- L16:2908 excites
- L14:374 suppresses
- L17:446 excites
- L2:1321 suppresses
- L7:631 excites
- L5:667 suppresses
- L10:576 suppresses
Feeds later layers
- L20:1476 suppresses
- L21:2623 suppresses
- L20:2613 suppresses
- L20:2426 suppresses
- L20:486 suppresses
- L20:272 suppresses
- L22:530 weakly suppresses
- L27:507 weakly excites
- L26:1539 weakly suppresses
- L27:535 weakly excites
- L20:824 weakly suppresses unclear: opposite pairs and citation digits
- L20:438 weakly excites
- L20:404 weakly suppresses
- L23:238 weakly excites
- L27:443 weakly excites
- L20:110 weakly suppresses
- L27:973 weakly excites
- L27:157 weakly excites
- L27:367 weakly excites
- L24:400 weakly excites
- L25:194 weakly suppresses
- L20:111 weakly suppresses
- L22:685 weakly suppresses
- L20:2553 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
- theirs
- _Syntax
- !_
- (Have
- пром
- зи
- Tehran
- 辘
- Garc
- ymes
Suppresses
- 事
- Pref
- 信
- ler
- <|endoftext|>
- 干
- 用
- 하
- 악
- bing