Neuron Explorer

Qwen3-0.6B

word endings split mid-word, like 'shower' in 'showered'

A neuron in layer 27 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the piece of a word that comes right before its ending, such as 'shower' in 'showered' or 'ith' in 'tithing'. Turning it up made the model drift into unrelated, loose text, and turning it down made it repeat 'ing' over and over.

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

    so many words, no right exists to have a machine gun. You can add the U.S. Court of Appeals for the Ninth Circuit to the list.⏎In affirming the conviction of Alaskan Matthew Wayne Henry for possessing a homemade machine gun Thursday

  2. #2 peak

    . Studies Cancer Fallout⏎- Lawrence Of Arabia's Hand-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 Dis

  3. #3 peak

    —one of the big shells struck the ground right in front of us and we all went up in the air. But we all come down again.”⏎York was showered with honors—the Distinguished Service Cross, the Croix de Guerre with

  4. #4 peak

    ’s tithes to Melchizedek. Do Abraham’s children not do the same?⏎From the examples of both Abraham and Jacob, we learn that tithing was not simply a commandment found in Moses’ law (Gen.

  5. #5 peak

    traditions of Christmas during slavery were tools for celebration in the harshest working and living conditions for blacks. While the whites in the “big house” were being showered with gifts and feast, they shared a portion of those with their captives

  6. #6 peak

    the ground only yards from the boundary fences. Sparks, rocks and soil hissed in blinding explosions in all direction. Those Israelites pressing against the barriers were showered with hot stones and sand. Gasping with terror, they stumbled and scrambled

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:inginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginginging

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

    making a shield, making a trebuchet, and more.⏎- Knowledge Box Central families: Use this alongside the Middle Ages lapbooks, foldables, and booklets.⏎- Tapestry of Grace families: Use for 9 weeks.⏎There

  2. #2 peak

    course that it’s the size of a freaking cat!⏎It has a silver coat of long, thick fur, making it not only the biggest, but also the cutest looking rat I have ever seen. This giant rat also seemed very tame and

  3. #3 peak

    City of Opelika, Alabama, had charged Jones, a member of the Jehovah’s Witnesses, with violating its licensing ordinance by selling books, operating as a book agent, and operating as a transient dealer, all without a license. The trial

  4. #4 peak

    so that she is certain not to go out of the lines while another little girl with the same princess coloring pages scribbles quickly in a hurry to get done first.⏎If you want to purchase such

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. L23:385 strongly suppresses
  2. L21:327 strongly suppresses
  3. L24:118 excites
  4. L5:849 suppresses
  5. L25:2123 suppresses
  6. L26:258 excites
  7. L18:784 excites
  8. L26:705 excites
  9. L26:729 excites
  10. L19:39 suppresses
  11. L19:910 excites
  12. L25:1493 excites
  13. L22:1654 suppresses
  14. L17:2183 suppresses
  15. L20:879 suppresses
  16. L20:819 suppresses
  17. L5:1278 excites
  18. L22:1141 suppresses
  19. L22:2769 suppresses
  20. L16:11 suppresses
  21. L26:685 suppresses
  22. L3:502 suppresses
  23. L19:1149 suppresses
  24. L19:233 suppresses

Feeds later layers

None: this is the last layer.

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.

4 values between -11.35 and -10.352 values between -10.35 and -9.352 values between -9.35 and -8.359 values between -8.35 and -7.3512 values between -7.35 and -6.3514 values between -6.35 and -5.3532 values between -5.35 and -4.3545 values between -4.35 and -3.3585 values between -3.35 and -2.35193 values between -2.35 and -1.35335 values between -1.35 and -0.351462 values between -0.35 and 0.651650 values between 0.65 and 1.6541 values between 1.65 and 2.653 values between 2.65 and 3.654 values between 3.65 and 4.651 values between 4.65 and 5.652 values between 5.65 and 6.652 values between 6.65 and 7.651 values between 8.65 and 9.641 values between 9.64 and 10.641 values between 10.64 and 11.641 values between 11.64 and 12.641 values between 14.64 and 15.641 values between 16.64 and 17.641 values between 19.64 and 20.641 values between 20.64 and 21.641 values between 24.64 and 25.641 values between 27.64 and 28.64offnegativepositiveoften

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. นโยบาย
  2. 失望
  3. 太久
  4. 主持召开
  5. 没想到
  6. 还以为
  7. 趣
  8. ours
  9. 有一次
  10. luận

Suppresses

  1. ing
  2. ed
  3. ING
  4. er
  5. ings
  6. ers
  7. ered
  8. ting
  9. edList
  10. edBy