Neuron Explorer

Qwen3-0.6B

the 'gro' fragment in words

A neuron in layer 26 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 small chunk 'gro' that begins words like 'grooves', 'groynes' and 'grocers'. Its negative side is weak and scattered over code and citation text, and steering gave mostly unrelated, sometimes broken 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

    support of the University Grants Commission, that 20 million cubic metres of granite waste had piled up in quarry belts of the State. The total requirement for the groynes was only 10 million cubic metres. By implementing the suggestion, the

  2. #2 peak

    on their doorposts, indicating a house as the home of a Jew.Read article on the web site⏎In post-Holocaust Europe you could often find the grooves or scars where mezuzahs had been removed or painted over during or

  3. #3 peak

    where they were, picked up their families, and moved to Allegany County. The towns of Richburg, Bolivar, Allentown, Gen<|endoftext|>A grooved pul

  4. #4 peak

    highly pressurised and finally breaks through the surface layer.⏎As the gas escapes, it carries the smaller, finer particles of the soil along with it, forming grooves under the ice. This "spider" effect indicates a spot where a

  5. #5 peak

    advantage. A rope block and tackle is usually for hand operation. To lift larger loads by hand, a chain is substituted for the rope and the pulleys have grooves for gripping the links. A differential pulley consists of two pulleys of

  6. #6 peak

    twin benefits, says expert⏎Plan will ensure protection of coastal areas, disposal of waste 20 million cubic metres of granite waste in quarry belts Total requirement for groynes 10 million cubic metres Make environmental audit mandatory, says top geologist

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:

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

    Oracle Solaris 11.1 Information Library|⏎- load exponent of a floating point number⏎c99 [ flag... ] file... -lm [ library... ] #include <math.h> double ldexp(double x, int exp

  2. #2 peak

    https://edis.ifas.ufl.edu (keyword search "business plan").⏎Prospective growers should compile

  3. #3 peak

    ry disease characteristics, and the clinical parameters.⏎The investigators scoured the Pubmed, EMBASE, and CINAHL databases from their inception to February of 20

  4. #4 peak

    withstand water submersion. Although “waterproof” does not have a technical definition, most companies who use the term have evaluated the product’s ingress protection (IP) rating in a laboratory (“ingress” means “entrance” and is

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. L22:878 strongly suppresses
  2. L21:946 strongly suppresses
  3. L25:311 strongly excites
  4. L25:161 suppresses
  5. L23:211 excites
  6. L19:763 suppresses
  7. L24:609 excites
  8. L19:364 suppresses
  9. L23:508 suppresses
  10. L24:1539 excites
  11. L25:112 excites
  12. L19:1308 suppresses
  13. L22:424 suppresses
  14. L23:777 excites
  15. L22:16 excites comma in dates and birth lines
  16. L19:1516 suppresses
  17. L18:1232 suppresses
  18. L25:1667 suppresses
  19. L19:674 suppresses
  20. L23:77 excites the space right before a number
  21. L18:246 suppresses
  22. L20:436 weakly excites
  23. L21:982 weakly excites
  24. L24:1758 weakly suppresses the 'Ar' syllable inside names

Feeds later layers

  1. L27:292 strongly suppresses
  2. L27:825 suppresses
  3. L27:1811 excites
  4. L27:1277 weakly suppresses
  5. L27:1867 weakly excites
  6. L27:731 weakly suppresses
  7. L27:1793 weakly suppresses
  8. L27:1294 weakly suppresses
  9. L27:904 weakly excites
  10. L27:2348 weakly excites
  11. L27:128 weakly suppresses unclear: opening quotes and closing brackets
  12. L27:286 weakly excites
  13. L27:503 weakly suppresses
  14. L27:958 weakly excites
  15. L27:1463 weakly suppresses
  16. L27:1386 weakly excites
  17. L27:297 weakly excites
  18. L27:744 weakly suppresses
  19. L27:170 weakly suppresses unclear: spaces in mixed-language text
  20. L27:2588 weakly excites
  21. L27:2245 weakly excites
  22. L27:2784 weakly excites
  23. L27:2262 weakly excites
  24. L27:2306 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 -9.38 and -9.052 values between -5.51 and -5.191 values between -4.87 and -4.541 values between -4.22 and -3.901 values between -3.26 and -2.931 values between -2.61 and -2.293 values between -2.29 and -1.971 values between -1.97 and -1.654 values between -1.65 and -1.323 values between -1.32 and -1.0013 values between -1.00 and -0.6881 values between -0.68 and -0.36515 values between -0.36 and -0.041985 values between -0.04 and 0.29832 values between 0.29 and 0.61285 values between 0.61 and 0.93117 values between 0.93 and 1.2542 values between 1.25 and 1.5815 values between 1.58 and 1.904 values between 1.90 and 2.221 values between 3.19 and 3.51offnegativepositiveoften

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. }
  2. }⏎
  3. }.⏎
  4. })}⏎
  5. }/>⏎
  6. }> ⏎
  7. }|
  8. }>⏎
  9. }));⏎
  10. }))⏎

Suppresses

  1. )
  2. ),
  3. ).
  4. ):
  5. );⏎
  6. ); ⏎
  7. );
  8. );⏎⏎
  9. )⏎⏎
  10. ); ⏎ ⏎