Neuron Explorer

Qwen3-0.6B

Chinese commas and list separators

A neuron in layer 15 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on commas in Chinese text (the Chinese comma and list separator), and sometimes on English commas that come before a second clause. It is a comma detector, mostly in Chinese sentences. Turning it up made the model write in a loose, rambling style.

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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

  1. #1 peak

    Pre-test)”的已达到83.3%,说明对比分析的研究成果对泰国学生学习汉语助动词具有较大的帮助,学生能更容易、更好地把握汉语助动词的语义、语法功能与特点。aux

  2. #2 peak

    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偏误。本文通过调查法、统计

  3. #3 peak

    .⏎Cleaner uniforms must be equipped with appropriate personal protective equipment, such as gloves, water shoes, dust masks, etc., which can not only improve work efficiency, but also reduce or avoid occupational hazards for cleaners.⏎It can avoid the breeding and

  4. #4 peak

    Fire tube boilers — Biomass boilers; Bubbling fluidized bed boilers; Circulating fluidized bed boilers⏎- Front<|endoftext|>CNY - 中國元,人民幣⏎最常用 CNY 交叉匯率⏎為什麼您

  5. #5 peak

    capitalist society, all human relationships are voluntary. Men are free to cooperate or not, to deal with one another or not, as their own individual judgments, convictions, and interests dictate. They can deal with one another only in terms of<|endoftext|>School

  6. #6 peak

    the software of the servo system hardware. There is a greater improvement. The drive components have evolved from the early thyristor SCR, high power transistor GTR, etc. to the current smart power device IPM.⏎In the analog AC servo

Steering

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.

turned down a lot
Prompt

Here is something I have been thinking about lately: "I know, I know, " the "s, " " (2145, 2145, 2145, 2145, 2145, 214

Most negative on

Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.

  1. #1 peak

    221BC). Unlike most Chinese idioms, 毋忘在莒 strangely has actually two stories of origin, which gave this idiom two

  2. #2 peak

    buried alive. This phobia peaked in the 18th and 19th centuries when epidemics of yellow fever and cholera commonly killed thousands and thousands of people. This fear was not totally unfounded. An Englishman, William

  3. #3 peak

    .⏎“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⏎Pin Yin Version Chun mian bu jue xiao Chu chu wen ti

  4. #4 peak

    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

Connections

Show on map

Which neurons in earlier layers feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.

Fed by earlier layers

  1. L10:155 strongly excites
  2. L13:44 strongly suppresses inside web addresses and citations
  3. L11:223 excites
  4. L2:24 suppresses
  5. L7:92 excites
  6. L4:76 excites
  7. L8:328 excites
  8. L3:130 excites
  9. L5:7 suppresses
  10. L10:370 suppresses
  11. L4:342 suppresses
  12. L12:158 suppresses
  13. L1:699 excites
  14. L9:17 suppresses
  15. L1:25 excites
  16. L10:669 suppresses
  17. L9:74 excites
  18. L11:840 suppresses
  19. L4:92 suppresses
  20. L9:207 weakly suppresses
  21. L4:42 weakly suppresses
  22. L7:72 weakly excites
  23. L10:1289 weakly suppresses
  24. L10:873 weakly excites

Feeds later layers

  1. L17:740 excites
  2. L18:1103 excites
  3. L16:995 excites
  4. L18:126 suppresses comma joining a full clause
  5. L17:472 weakly excites
  6. L16:797 weakly suppresses
  7. L25:2797 weakly excites
  8. L18:861 weakly excites
  9. L17:1756 weakly excites
  10. L17:1772 weakly excites
  11. L20:822 weakly excites
  12. L22:654 weakly excites
  13. L26:25 weakly suppresses
  14. L16:28 weakly suppresses
  15. L22:669 weakly excites
  16. L18:276 weakly suppresses
  17. L17:2008 weakly excites
  18. L19:1966 weakly excites
  19. L18:1842 weakly excites
  20. L19:2759 weakly excites
  21. L16:1142 weakly excites
  22. L27:2038 weakly excites
  23. L21:33 weakly suppresses
  24. L18:902 weakly excites

Activation histogram

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.

2 values between -1.44 and -1.344 values between -1.34 and -1.244 values between -1.24 and -1.146 values between -1.14 and -1.056 values between -1.05 and -0.954 values between -0.95 and -0.8516 values between -0.85 and -0.7517 values between -0.75 and -0.6517 values between -0.65 and -0.5526 values between -0.55 and -0.4539 values between -0.45 and -0.3578 values between -0.35 and -0.26146 values between -0.26 and -0.16399 values between -0.16 and -0.06962 values between -0.06 and 0.04902 values between 0.04 and 0.14656 values between 0.14 and 0.24371 values between 0.24 and 0.34146 values between 0.34 and 0.4452 values between 0.44 and 0.5319 values between 0.53 and 0.6310 values between 0.63 and 0.739 values between 0.73 and 0.835 values between 0.83 and 0.932 values between 0.93 and 1.034 values between 1.13 and 1.232 values between 1.23 and 1.321 values between 1.32 and 1.422 values between 1.52 and 1.621 values between 2.41 and 2.51offnegativepositiveoften

Logit lens

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 layers instead.

Promotes

  1. iversary
  2. 一回事
  3. 他们在
  4. úa
  5. arded
  6. heels
  7. rove
  8. 如果是
  9. 淹没
  10. emente

Suppresses

  1. ,
  2. ред
  3. (⏎
  4. comm
  5. [⏎
  6. (
  7. �
  8. Agent
  9. migration
  10. saying