Neuron Explorer

Qwen3-0.6B

award and prize names

A neuron in layer 5 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron lights up on the word right after or around the names of awards and prizes, like 'Awards', 'Prize', 'Nobel' or 'Grammy'. Turning it up made the model's writing drift into loose, unrelated sentences rather than anything about prizes.

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

    :⏎Get started on your Next Big Thing Nova award with these ideas and related achievements:⏎STEM stands for science, technology, engineering, and mathematics. The NOVA Awards program combines knowledge of STEM concepts with hands on activities.<|endoftext|>The sheer amount of

  2. #2 peak

    in the American West, 1850-1920⏎Distinguished Book Award, Conference for the History of American Religious⏎2013 Armitage

  3. #3 peak

    technology, all farmers could produce enough to feed their families.⏎Gomel, aged 37, has been selected as an Associate Laureate in the Rolex Awards for an ambitious project to encourage more than 500 families in the areas

  4. #4 peak

    it Ravi Shankar spread Indian music in Europe specially Paris and got grammy in 1967 and 1972 (“Instrumentalists

  5. #5 peak

    , to recognize exceptional impact through an annual WISE awards program, and to promote innovation and educational engagement on a global scale.⏎The summit

  6. #6 peak

    mothers ruled the world, there wouldn’t be any . . . wars in the first place.”⏎SALLY FIELD, acceptance speech at 2007 Emmy Awards⏎What differences do you see in countries who have endured war compared to those who

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 (5) is provided the data, and they should be the all individuals data is, Ministry can be in that the initial information is not based for the data is are rely on the the information is (1 may be the data.

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

    On the 83rd anniversary of a West Chicago Charity Minstrel Show, the high school’s Glee Club will perform a twoact program of songs, dances, and skits performed in blackface and featuring the popular

  2. #2 peak

    in 95 events. The U.S. hosted the Olympics for the first time in 1904, in St. Louis. The third Olympiad marked the first time gold, silver and bronze medals were given out to first,

  3. #3 peak

    an event that Jewish people all around the world will be celebrating on the upcoming holiday of Shavuos, the Festival of Weeks. [This year Shavuos begins on Saturday evening May 23

  4. #4 peak

    , a semen like substance issues from the white berries. Mistletoe was considered a magickal aphrodisiac. Girls standing under a sprig of mistletoe were asking for a bit more than a kiss, it seems

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. L4:1363 excites
  2. L4:2220 excites
  3. L4:2390 weakly suppresses
  4. L4:2954 weakly suppresses
  5. L4:2362 weakly suppresses
  6. L3:1047 weakly suppresses
  7. L3:189 weakly excites
  8. L4:2149 weakly excites
  9. L4:1369 weakly excites
  10. L4:2350 weakly excites
  11. L3:2644 weakly suppresses
  12. L3:3042 weakly suppresses
  13. L3:2306 weakly suppresses
  14. L4:2680 weakly suppresses
  15. L3:836 weakly excites
  16. L4:2541 weakly excites
  17. L4:1565 weakly suppresses
  18. L2:111 weakly suppresses
  19. L0:149 weakly suppresses
  20. L3:1112 weakly suppresses
  21. L2:493 weakly suppresses
  22. L4:2021 weakly excites
  23. L0:2882 weakly excites
  24. L0:30 weakly suppresses

Feeds later layers

  1. L7:2618 excites
  2. L9:647 excites
  3. L9:805 excites
  4. L18:537 suppresses
  5. L20:1909 excites
  6. L9:2099 excites
  7. L8:1499 weakly excites
  8. L12:582 weakly excites
  9. L17:1203 weakly excites
  10. L7:721 weakly excites
  11. L17:2415 weakly suppresses
  12. L21:1850 weakly excites
  13. L12:559 weakly excites
  14. L6:380 weakly excites
  15. L7:1321 weakly excites
  16. L14:2028 weakly suppresses the '1' of numbers like 10
  17. L24:2709 weakly excites
  18. L11:3047 weakly suppresses
  19. L17:1293 weakly excites
  20. L20:2500 weakly suppresses
  21. L10:2222 weakly suppresses
  22. L22:818 weakly excites
  23. L11:2280 weakly suppresses
  24. L12:1075 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.

3 values between -0.72 and -0.5810 values between -0.58 and -0.4368 values between -0.43 and -0.28581 values between -0.28 and -0.131906 values between -0.13 and 0.021085 values between 0.02 and 0.16207 values between 0.16 and 0.3124 values between 0.31 and 0.4611 values between 0.46 and 0.613 values between 0.61 and 0.762 values between 0.90 and 1.051 values between 1.05 and 1.202 values between 1.35 and 1.491 values between 1.79 and 1.941 values between 1.94 and 2.081 values between 2.53 and 2.681 values between 3.27 and 3.411 values between 5.04 and 5.19offnegativepositiveoften

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. awards
  2. 评选
  3. tees
  4. Outstanding
  5. Awards
  6. -winning
  7. qualify
  8. awarded
  9. award
  10. 下乡

Suppresses

  1. chte
  2. prox
  3. ầm
  4. чат
  5. 教学质量
  6. Bre
  7. }%
  8. chr
  9. README
  10. _RANGE