Neuron Explorer

Qwen3-0.6B

unclear: first pieces of split words

A neuron in layer 23 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

It seems to switch on at the start of a word that gets cut into pieces, such as 'Mo' in Moist or 'Wood' in Woodrow. The pattern is loose and the turned-up and turned-down outputs did not show a clear theme.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    rescens (T.& G.) A. Gray. Nutlets slightly wrinkled or pebbly; plants 0.5-3 m tall⏎Moist to mesic meadows, ditches, streambanks, grasslands,

  2. #2 peak

    the first permanent settlement in Carolina. They then divided the new county into four precincts: Currituck , Perquimans , Pasquotank , and Chowan . Albemarle County was subsequently enlarged, and in 16

  3. #3 peak

    practice the alphabet as they travel from letter to letter in the maze. An open area at the front with toddler<|endoftext|>Mental Health and Wellbeing for Students⏎Moat gives students a platform in which to

  4. #4 peak

    for design projects. Projects are developed from concepts learned in Wood and Metal Design, Interior Design and/or Intro to Architecture. New technologies such as

  5. #5 peak

    sources to change U.S. law. Learn what you can do about bad decisions.<|endoftext|>8.75" by 5.5"⏎Wood, Paper, Leather⏎19th Century Culture; Religious Beliefs; Native

  6. #6 peak

    often."⏎At the beginning of the program, 41 percent of students reported experiencing bullying, while 34 percent did so after the two-month program.⏎Moises Prospero, a Utah research consultant on criminal and social justice issues with a

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: if a person is able to create a final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final final

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

    ⏎horizontally for 1 - 5 mm before bending up or curling downward. This⏎curling tendency produces the characteristic tangled branches. Branches⏎coarser below (1.5 mm diam.) more slender above (0.2

  2. #2 peak

    ama (a tuber also known as sengkwang, yam bean, singkamas, Mexican turnip), amaranth (also called pigweeds) and mucuma (a tropical legume). “Milpa crops are

  3. #3 peak

    acids in the dough, and the final, hottest process called caramelization. But watch the video to get the whole, fun story!⏎Dig Deeper, Get Sweeter⏎As with all TED ED lessons, there are links for those with a

  4. #4 peak

    word meaning simply “box.” This ark, on the other hand, is represented by a Hebrew word meaning “chest” or “coffin.” It was first used in Genesis 50:26,

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. L16:74 excites
  2. L17:840 suppresses
  3. L20:417 excites
  4. L7:2731 suppresses
  5. L18:288 suppresses
  6. L2:2029 excites
  7. L14:246 excites
  8. L3:219 suppresses
  9. L18:880 excites
  10. L20:1442 excites
  11. L0:2471 excites
  12. L5:340 excites zeros inside long numbers
  13. L19:754 excites
  14. L18:569 excites
  15. L21:463 suppresses
  16. L20:430 suppresses
  17. L21:334 weakly excites
  18. L10:1228 weakly suppresses
  19. L19:1801 weakly suppresses
  20. L2:2053 weakly suppresses
  21. L18:533 weakly suppresses
  22. L18:541 weakly excites
  23. L20:416 weakly suppresses the "1" in years 191x
  24. L18:1442 weakly suppresses unclear: word fragments and digits inside numbers

Feeds later layers

  1. L24:408 suppresses
  2. L25:2519 suppresses
  3. L27:763 excites
  4. L26:739 suppresses
  5. L26:976 excites the '1' starting a century
  6. L24:2463 excites
  7. L26:905 excites
  8. L27:391 suppresses
  9. L25:147 excites
  10. L24:800 suppresses colon in clock times
  11. L24:278 excites
  12. L26:724 weakly suppresses
  13. L27:1178 weakly excites
  14. L27:1100 weakly excites
  15. L25:1613 weakly excites
  16. L26:294 weakly excites
  17. L26:2396 weakly excites
  18. L25:342 weakly suppresses
  19. L26:102 weakly excites
  20. L25:2029 weakly excites
  21. L25:58 weakly suppresses colon in clock times
  22. L27:678 weakly suppresses
  23. L27:785 weakly suppresses
  24. L24:1253 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 -30.48 and -29.431 values between -25.19 and -24.131 values between -24.13 and -23.081 values between -12.49 and -11.431 values between -10.38 and -9.322 values between -9.32 and -8.263 values between -8.26 and -7.203 values between -7.20 and -6.145 values between -6.14 and -5.082 values between -5.08 and -4.024 values between -4.02 and -2.979 values between -2.97 and -1.9126 values between -1.91 and -0.853523 values between -0.85 and 0.21290 values between 0.21 and 1.2715 values between 1.27 and 2.335 values between 2.33 and 3.386 values between 3.38 and 4.441 values between 4.44 and 5.503 values between 5.50 and 6.563 values between 6.56 and 7.621 values between 7.62 and 8.681 values between 8.68 and 9.741 values between 10.79 and 11.85offnegativepositiveoften

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. тур
  3. ügen
  4. 八大
  5. ChangedEventArgs
  6. culate
  7. vro
  8. bate
  9. 耗
  10. adratic

Suppresses

  1. izzo
  2. isper
  3. ales
  4. ят
  5. izzato
  6. operator
  7. ourke
  8. 第四
  9. fourth
  10. 四是