Neuron Explorer

Qwen3-0.6B

blank before an expected number

A neuron in layer 22 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 responds to an empty spot in the text where a number is about to come, such as after '3 x' in a physics example or after a date. The same kind of spot also appears on the negative side, so the sign seems to matter in ways that are hard to read. Turning it up pushed the model toward writing numbers and digit patterns.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    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

  2. #2 peak

    年7月6日 at 9:45 AMに Megan Smith さんが最後の更新を行いました。⏎2020年7月6日 9:45 AM #2107

  3. #3 peak

    ) the successful deflection of a shot by touching part of the ball on its way to the basket, thereby preventing a field goal⏎Một loại quần short ống rộng vừa vặn, dệt kim hoặc quần soóc nam thường

  4. #4 peak

    �� : sleep 不 : not 觉 : to feel, wake up 晓 : dawn⏎Meaning for Lien 2 处 : place 闻 : hear 啼 : cry 茑 : bird⏎Meaning for

  5. #5 peak

    Influenza virus activity in the world⏎12 August 2011⏎During weeks 29 to 30 (17 July

  6. #6 peak

    a substance of one density into a substance of different density, the speed of light changes and bending of light occurs.⏎In air, light travels at 3 x 108 m/sec but in water, it travels only about 2.

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: the fact that the fact that the 3000 (in the fact that the fact that the fact ( 3000 ) ) 3000 ( 30

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

    cookies or want to be informed before they are stored on their computer can configure their browser for this purpose.⏎Most of today’s browsers allow the management of cookies in 3 different ways:⏎The cookies are never accepted.⏎For each cookie, the browser

  2. #2 peak

    The answer to soaring gas prices lies in renewable heating, a technology that is ready to replace gas boilers but only affordable in 8 EU countries.⏎To enable consumers to overcome the upfront cost for switching to clean

  3. #3 peak

    segregationists, whatever one thought of Rosa Parks, her sit-down in 1955 in Montgomery, Alabama, was the ultimate moment of truth in a 350-year-long crisis and everybody knew it. What she did demanded a

  4. #4 peak

    the atheists are about this report. They are positively gloating.⏎In 1988, Richard Lenski, Michigan State University, East Lansing, founded 12 cultures of E.

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. L19:581 strongly suppresses blank slot before a number
  2. L18:377 excites number right after 'about' or 'less than'
  3. L20:344 excites hyphen inside compound words
  4. L4:53 excites blank slot before a number
  5. L20:265 excites
  6. L21:431 suppresses
  7. L21:339 excites spot just before a year
  8. L18:375 suppresses
  9. L17:402 suppresses
  10. L16:471 suppresses blank spot before a year
  11. L20:749 excites
  12. L15:307 suppresses
  13. L19:89 excites
  14. L13:923 suppresses word right before a number or year
  15. L19:165 excites
  16. L16:0 suppresses
  17. L21:173 excites
  18. L16:363 suppresses
  19. L1:267 suppresses
  20. L17:285 suppresses
  21. L20:629 excites
  22. L5:104 excites
  23. L16:406 excites
  24. L1:217 suppresses

Feeds later layers

  1. L27:170 strongly excites unclear: spaces in mixed-language text
  2. L23:77 excites the space right before a number
  3. L26:97 suppresses
  4. L25:33 excites spot where a number is due
  5. L26:423 suppresses
  6. L26:103 excites blank space before a number
  7. L26:624 suppresses
  8. L27:283 excites the spot right before a number
  9. L26:172 excites
  10. L24:33 suppresses
  11. L27:97 weakly excites
  12. L27:33 weakly excites
  13. L23:2450 weakly suppresses
  14. L27:220 weakly excites end-of-text token between documents
  15. L26:254 weakly suppresses
  16. L25:222 weakly suppresses
  17. L27:413 weakly excites
  18. L24:1223 weakly suppresses
  19. L25:232 weakly suppresses
  20. L23:80 weakly excites blank slot before a year or number
  21. L26:145 weakly excites
  22. L26:611 weakly suppresses
  23. L27:2757 weakly suppresses
  24. L25:359 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.

1 values between -20.66 and -19.351 values between -10.23 and -8.932 values between -6.33 and -5.021 values between -5.02 and -3.724 values between -2.42 and -1.123869 values between -1.12 and 0.1911 values between 0.19 and 1.491 values between 1.49 and 2.791 values between 4.10 and 5.401 values between 5.40 and 6.701 values between 6.70 and 8.002 values between 8.00 and 9.311 values between 10.61 and 11.911 values between 11.91 and 13.222 values between 13.22 and 14.521 values between 14.52 and 15.824 values between 15.82 and 17.122 values between 17.12 and 18.431 values between 28.85 and 30.151 values between 30.15 and 31.45offnegativepositiveoften

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. že
  2. ẩm
  3. -arm
  4. 里的
  5. ы
  6. -fluid
  7. 侵蚀
  8. -ing
  9. 诃
  10. 亲

Suppresses

  1. out
  2. fall
  3. PUBLIC
  4. rab
  5. ?><?
  6. law
  7. 气
  8. público
  9. run
  10. mux