Neuron Explorer

Qwen3-0.6B

the space right before a number

A neuron in layer 23 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 the blank space that comes just before a number, such as the gap before 14 in 'ages of 14 and 16'. Turning it up made the model's output fall into repeated filler words, not number-related text.

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

    -Drawn, WWI Map Is Up for Auction⏎- Thousands Of FBI Documents About Civil Rights Era Destroyed By Flooding⏎- Ancient Egyptian Woman with 70 Hair Extensions Discovered⏎- Europeans drawn from three ancient 'tribes

  2. #2 peak

    iers and the key to understanding why mass exists. The most recent book on the subject, The God Particle, was a bestseller.⏎Students in your classes between 7 and 16 years of age can<|endoftext|>Harivamsha MahA

  3. #3 peak

    3/74 that ended the "economic miracle" once and for all. By the time Herbert Grünewald succeeded Kurt Hansen as Management Board Chairman following the 1974 Annual Stockholders' Meeting, the global economy was undergoing a

  4. #4 peak

    chaos. But there are also peaceful applications<|endoftext|>An estimated 50,000 to 60,000 farmworkers between the ages of 14 and 16 might be affected by proposed changes to child labor laws

  5. #5 peak

    changes, the Department of Labor estimates that between 50,000 and 60,000 hired farmworkers are between the ages of 14 and 16. The most important of the proposed child labor changes

  6. #6 peak

    Course Description and Rationale:⏎A 2006 report by the Corporation for National and Community Service found that 3.3 million college students volunteered in 2005. This course recognizes college students as volunteer leaders with the abilities

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 "

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

    0s, but if nobody recorded it in publication<|endoftext|>Opportunities and Challenges in High Pressure Processing of Foods⏎By Rastogi, N K; Raghavarao, K S M S; Balasubramaniam, V

  2. #2 peak

    required for flight, puncturing a lung, and barely missing his heart.⏎"The pellet broke the coracoid bone," noted Cape Wildlife Center veterinarian Dr. Roberto Aguilar. "It is difficult to set a broken shoulder bone because of the challenge

  3. #3 peak

    2017 WHSmith and its suppliers.⏎WHSmith High Street Limited Greenbridge Road, Swindon, Wiltshire, United Kingdom, SN3 3

  4. #4 peak

    successful, the operation could also generate millions of dollars to be used to combat the shrinking of the Salton Sea and the<|endoftext|>The Art of Story Writing⏎by Nathaniel C. Fowler⏎Publisher: Sulley and Kleinteich 19

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. L18:106 suppresses
  3. L21:339 excites spot just before a year
  4. L22:372 excites empty spot before a number or "3D"
  5. L22:824 excites blank before an expected number
  6. L4:53 excites blank slot before a number
  7. L21:173 excites
  8. L18:377 excites number right after 'about' or 'less than'
  9. L14:191 excites blank spot before a run-on number
  10. L17:285 suppresses
  11. L20:2446 suppresses
  12. L20:433 excites
  13. L19:781 excites
  14. L16:41 suppresses
  15. L6:0 excites
  16. L19:1207 weakly excites
  17. L9:7 weakly excites
  18. L8:111 weakly suppresses the space before a year in dates
  19. L16:208 weakly suppresses
  20. L5:334 weakly suppresses
  21. L22:216 weakly excites
  22. L20:344 weakly excites hyphen inside compound words
  23. L16:91 weakly excites
  24. L18:284 weakly suppresses

Feeds later layers

  1. L27:170 strongly excites unclear: spaces in mixed-language text
  2. L25:222 strongly suppresses
  3. L27:71 excites
  4. L27:237 suppresses
  5. L26:17 suppresses
  6. L27:283 excites the spot right before a number
  7. L25:33 excites spot where a number is due
  8. L26:396 suppresses
  9. L26:423 suppresses
  10. L25:239 suppresses
  11. L25:161 excites
  12. L26:542 suppresses
  13. L25:736 excites
  14. L26:97 suppresses
  15. L27:83 suppresses
  16. L27:63 excites closing parenthesis in data lists
  17. L27:112 suppresses
  18. L27:65 suppresses
  19. L25:311 excites
  20. L25:112 suppresses
  21. L27:22 suppresses
  22. L27:825 excites
  23. L27:292 excites
  24. L27:1017 suppresses

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.

122 values between -3.71 and -1.982478 values between -1.98 and -0.261256 values between -0.26 and 1.4713 values between 1.47 and 3.196 values between 3.19 and 4.912 values between 4.91 and 6.641 values between 6.64 and 8.361 values between 8.36 and 10.092 values between 11.81 and 13.531 values between 16.98 and 18.711 values between 20.43 and 22.151 values between 25.60 and 27.332 values between 27.33 and 29.051 values between 29.05 and 30.771 values between 32.50 and 34.221 values between 35.94 and 37.673 values between 39.39 and 41.121 values between 41.12 and 42.842 values between 42.84 and 44.563 values between 49.74 and 51.461 values between 51.46 and 53.185 values between 53.18 and 54.911 values between 56.63 and 58.351 values between 61.80 and 63.532 values between 63.53 and 65.25offnegativepositiveoften

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. }>
  2. }>⏎
  3. themselves
  4. #
  5. //
  6. Movement
  7. )[
  8. ):⏎
  9. }(
  10. isible

Suppresses

  1. ⏎
  2. ⏎
  3. ⏎
  4. ⏎