Neuron Explorer

Qwen3-0.6B

square bracket opening an inserted note

A neuron in layer 24 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 a square bracket that opens an added note, citation or translation, such as the '[' in a reference or a clarification inside a sentence. It is weaker on brackets inside code or list indices. Turning it up or down produced rambling text that does not clearly relate to brackets.

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

    ians Institute Conference, held at the University of Virginia on October 1-3, 2010. The conference had the theme "By the Numbers." Attended by "analog" Victorianists as well as some budding digital human

  2. #2 peak

    , writing for the majority, slapped down New York’s 100-year-old concealed carry licensing scheme Thursday on the grounds that it has no historical analogue. [Wait, doesn’t a law that’s 100 years old have some

  3. #3 peak

    ave, Antoine Pierre Joseph Marie (äNtwänˈ pyĕr zhōzĕfˈ märēˈ bärnävˈ) [key], 1761–93, French revolutionary. A

  4. #4 peak

    ogdian language, this name could not be transformed from a word vihara for which in the Sogdian used the word βrx’r [Lurje, 2004, p. 20].

  5. #5 peak

    better known dwarf galaxies.⏎The new galaxies [image] were detected over the past two years as part of the Sloan Digital Sky

  6. #6 peak

    . Please check the complete study for more information on the data and analysis.<|endoftext|>Parus niger (Southern black tit)⏎Gewone swartmees [Afrikaans]; Isicukujeje [Xhosa];⏎Xid

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: those who go from those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those those

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

    pwγ’r) and Puxar (pwx’r) [Smirnova, 1981, p. 34 (792 – 796); Smirnova, 198

  2. #2 peak

    2, 3, 4, 5] i = len(given_arr) - 1 while(i >= 0): print(given_arr[i]) i = i - 1⏎In this program,⏎- given_arr is

  3. #3 peak

    of environmental, economic, and social sustainability is predicated on changing human behavior; the purview of Psychologists (reviewed in , see also [2,3,4,5,6,7]).⏎So-called “environmental problems”

  4. #4 peak

    MAS[1:0]=01), and A[1:0] for ARM instruction fetches (nOPC=0 and MAS[1:0]=10).⏎Article last edited on: 2008-0

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. L23:353 strongly excites
  2. L19:1048 excites
  3. L21:458 excites
  4. L21:3039 suppresses
  5. L19:1581 suppresses
  6. L22:275 excites
  7. L18:866 suppresses
  8. L18:1375 weakly suppresses
  9. L23:1646 weakly excites
  10. L3:2965 weakly suppresses
  11. L21:1269 weakly suppresses "the" in "on the other hand"
  12. L17:1434 weakly excites
  13. L20:2391 weakly excites
  14. L20:2336 weakly excites
  15. L22:716 weakly suppresses
  16. L13:2281 weakly suppresses
  17. L23:2829 weakly excites
  18. L14:1801 weakly excites
  19. L6:3027 weakly suppresses
  20. L22:118 weakly suppresses
  21. L3:1078 weakly suppresses
  22. L2:2416 weakly excites
  23. L7:1461 weakly suppresses
  24. L13:113 weakly excites

Feeds later layers

  1. L27:1421 strongly suppresses
  2. L25:925 strongly suppresses
  3. L26:814 strongly suppresses
  4. L27:1363 strongly suppresses
  5. L26:936 strongly suppresses
  6. L27:1083 suppresses
  7. L26:557 suppresses
  8. L27:2612 suppresses
  9. L27:1530 excites
  10. L26:1320 excites
  11. L26:275 suppresses
  12. L27:697 weakly excites
  13. L27:877 weakly suppresses
  14. L25:192 weakly excites
  15. L25:213 weakly suppresses
  16. L27:2870 weakly suppresses
  17. L26:846 weakly excites
  18. L25:1491 weakly excites
  19. L27:197 weakly excites
  20. L27:2966 weakly suppresses
  21. L26:208 weakly suppresses
  22. L25:772 weakly excites
  23. L26:2116 weakly excites
  24. L26:730 weakly 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.

1 values between -24.39 and -23.261 values between -21.00 and -19.871 values between -10.83 and -9.703 values between -8.57 and -7.444 values between -7.44 and -6.311 values between -6.31 and -5.183 values between -5.18 and -4.056 values between -4.05 and -2.9218 values between -2.92 and -1.7958 values between -1.79 and -0.662962 values between -0.66 and 0.47838 values between 0.47 and 1.6010 values between 1.60 and 2.731 values between 15.16 and 16.291 values between 19.68 and 20.81offnegativepositiveoften

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. '}}⏎
  4. "}}⏎
  5. "}},⏎
  6. "},"
  7. }}
  8. '))⏎
  9. ))}⏎
  10. }}⏎

Suppresses

  1. ]
  2. "]
  3. 】
  4. ]⏎
  5. ]
  6. `]
  7. !]
  8. +]
  9. "]⏎
  10. ']