Neuron Explorer

Qwen3-0.6B

comma inside a stacked adjective list

A neuron in layer 2 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 comma that sits between two adjectives describing the same noun, like "a broad, flat tail" or "a sturdy, uncluttered surface". It does not react to the adjectives themselves. Turning it up did not produce a clear change that matches this.

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

    Space Agency. The High Resolution Stereo Camera, or HRSC, on board the orbiter is producing 3-D images of the planet's surface.⏎These sharp, panoramic, full-color pictures provided fodder for a third Nature report. In it,

  2. #2 peak

    Renewables Alliance (NARA) Education group aims to:⏎- Meet the workforce needs of the bio-energy/bioproducts economy⏎- Develop a broad, integrated view of the biofuels problem among emerging scientists and engineers⏎- Enh

  3. #3 peak

    that ultimately reintroduced the threatened species back into the wild. The zoologists used genetic information taken from the remaining black-footed ferrets, and bred a strong, biodiverse population that could keep the animals healthy and, more importantly, increase numbers

  4. #4 peak

    more than 3 feet in length, including the tail. These semi-aquatic mammals have webbed hind feet, large incisor teeth, and a broad, flat tail.⏎Once among the most widely distributed mammals in North America, beavers

  5. #5 peak

    thin horses forward, and the miserable animals, cold and hungry, struggled to pull the overfilled cars through the slippery streets.⏎Suddenly Henry Bergh, a tall, handsome gentleman in a black silk hat, stood on the tracks in front of the

  6. #6 peak

    , it didn’t. In 2007, Google engineers asked public-transit agencies across the country to submit their arrival and departure data in a simple, standard, open format—a text file, basically, with a bunch of numbers separated

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 am a fan of the show, and I have a love for the show" So, the show is the love you, and the show is a love you So, the show is the love you So, the show

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

    arian publication is a facsimile reprint of the unique. because of its age, it could comprise imperfections similar to marks, notations, marginalia and incorrect pages. simply because we think this paintings is culturally very important, we

  2. #2 peak

    Numbers decline thereafter and by November⏎any Warbling Vireo is rather scarce and late.⏎Warbling Vireo is a mid-sized, slim, small-billed vireo that is reasonably long-winged, long-tailed, and

  3. #3 peak

    that are late in filling out, or annuals to fill in any bare soil the bulb's dormant state leaves behind.⏎Sensitive to

  4. #4 peak

    .iii.87f.⏎2. Yodhājīva Sutta. On five kinds of warriors: those who go down into the thick of the fight where they are overpowered, those who are wounded and die on the

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. L1:861 suppresses
  2. L0:178 weakly suppresses
  3. L0:1301 weakly suppresses
  4. L0:1224 weakly suppresses
  5. L1:682 weakly excites
  6. L1:2579 weakly suppresses
  7. L0:1472 weakly excites
  8. L0:550 weakly suppresses
  9. L1:416 weakly suppresses
  10. L1:241 weakly suppresses
  11. L1:124 weakly suppresses
  12. L1:777 weakly excites
  13. L1:1612 weakly excites
  14. L1:1657 weakly excites
  15. L1:898 weakly excites
  16. L1:1464 weakly excites
  17. L0:97 weakly excites
  18. L1:1996 weakly excites
  19. L1:2010 weakly excites
  20. L1:2683 weakly excites
  21. L1:1312 weakly suppresses
  22. L1:560 weakly excites
  23. L1:2507 weakly suppresses
  24. L1:81 weakly suppresses

Feeds later layers

  1. L7:173 excites
  2. L3:58 excites
  3. L6:199 excites
  4. L4:128 excites
  5. L13:1315 excites
  6. L7:400 suppresses
  7. L18:1642 excites
  8. L17:390 excites
  9. L13:976 excites
  10. L23:324 weakly excites
  11. L14:2152 weakly excites
  12. L16:1015 weakly excites
  13. L18:1569 weakly excites
  14. L14:404 weakly excites
  15. L16:1819 weakly excites
  16. L20:383 weakly excites
  17. L19:2717 weakly excites
  18. L7:2102 weakly suppresses
  19. L7:1313 weakly excites
  20. L18:1129 weakly excites
  21. L4:951 weakly excites
  22. L14:1956 weakly excites
  23. L21:2239 weakly suppresses
  24. L4:132 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.

2 values between -1.47 and -1.343 values between -1.34 and -1.201 values between -1.20 and -1.071 values between -1.07 and -0.934 values between -0.93 and -0.803 values between -0.80 and -0.666 values between -0.66 and -0.5334 values between -0.53 and -0.3986 values between -0.39 and -0.26308 values between -0.26 and -0.131622 values between -0.13 and 0.011470 values between 0.01 and 0.14295 values between 0.14 and 0.2842 values between 0.28 and 0.419 values between 0.41 and 0.555 values between 0.55 and 0.683 values between 0.68 and 0.822 values between 0.82 and 0.951 values between 0.95 and 1.092 values between 1.22 and 1.363 values between 1.49 and 1.621 values between 2.16 and 2.301 values between 2.43 and 2.571 values between 2.97 and 3.101 values between 3.51 and 3.642 values between 3.78 and 3.91offnegativepositiveoften

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. 斑
  2. WSC
  3. щен
  4. etyl
  5. UED
  6. usses
  7. したもの
  8. основан
  9. asses
  10. awl

Suppresses

  1. expected
  2. 什么呢
  3. expected
  4. ��
  5. rob
  6. emploi
  7. !⏎⏎
  8. reason
  9. ramid
  10. apixel