Neuron Explorer

Qwen3-0.6B

the gap before a year or number

A neuron in layer 22 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 empty space that comes right before a year or number, such as the gap in "original 13 colonies" or "in 1776". Turning it down made the model loop on the word "school", and turning it up gave unrelated 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

    United States in 1777. Our nation's first official flag had 13 stars and 13 red and white stripes to represent the original 13 colonies that broke from British rule. The stars were arranged on a blue

  2. #2 peak

    1776 will find there is no getting around the special role New Jersey played in reigniting the American Revolution.⏎As one of the original 13 colonies, N.J. responded forcefully to the phrase "no taxation without

  3. #3 peak

    the case may be).⏎The United States has operated under two constitutions. The first, The Articles of Confederation, was in effect from March 1, 1781, when Maryland ratified it. The second, The Constitution,

  4. #4 peak

    Security (DHS)’s Customs and Border Protection apprehended 101,952 unaccompanied children (aliens under the age of 18 who either came to the United States without authorization or overstayed their visa

  5. #5 peak

    history, the U.S. Civil War, was resolved by amendments—creating the legal foundation for the equal rights of African Americans. The 13th, 14th and 15h Amendments—collectively known as the Civil

  6. #6 peak

    than 70 years for women to receive the right to vote, a goal that finally was achieved when the Nineteenth Amendment passed in 1920. Three years later, at the 75th anniversary

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: school school school, school school school, school school school, school school school, school school school, school school school, school school school, school school school, school school school, school school school, school school school, school school school,

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

    for multiplication.⏎When should kids memorize times tables? When their love for multiplication is strong enough.<|endoftext|>What Is Sufism?⏎Sufism, mystical Islamic belief and practice in which Muslims seek to find the truth of divine love and knowledge

  2. #2 peak

    Taoism is the belief in balancing forces, or yin and yang.⏎Where is Taoism mostly practiced?⏎Today, Taoism is recognized as one of the great world religions and continues to be practiced by people in China and throughout the world.⏎

  3. #3 peak

    Islamic parties - most prominently the Barelvi school of Islam.⏎Pakistan's laws became particularly severe between 1980 and

  4. #4 peak

    practice within the ambit of Shia Islam during the late medieval period. Although Sufis were opposed to dry legalism, they strictly observed Islamic law and belonged to various schools of Islamic jurisprudence and theology.⏎Sufis have been characterized by

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. L6:2765 excites
  2. L5:2495 weakly excites
  3. L2:879 weakly suppresses
  4. L20:1341 weakly excites
  5. L7:2681 weakly excites
  6. L20:139 weakly suppresses
  7. L11:2071 weakly suppresses
  8. L4:2191 weakly excites
  9. L20:2768 weakly suppresses
  10. L18:2650 weakly excites
  11. L16:1377 weakly excites
  12. L20:1497 weakly excites
  13. L14:797 weakly excites
  14. L18:2555 weakly excites
  15. L8:2511 weakly excites
  16. L6:2814 weakly excites
  17. L10:912 weakly excites
  18. L11:1851 weakly excites
  19. L1:1675 weakly suppresses
  20. L2:444 weakly suppresses
  21. L20:1130 weakly excites
  22. L14:522 weakly suppresses
  23. L19:2617 weakly excites
  24. L4:1793 weakly suppresses

Feeds later layers

  1. L27:2614 suppresses
  2. L26:2859 weakly suppresses
  3. L26:2078 weakly suppresses
  4. L23:1904 weakly excites
  5. L27:868 weakly suppresses
  6. L25:2512 weakly suppresses
  7. L24:2561 weakly suppresses
  8. L27:2398 weakly suppresses
  9. L24:1980 weakly excites
  10. L26:2499 weakly suppresses
  11. L25:1540 weakly suppresses
  12. L27:1379 weakly suppresses
  13. L24:2205 weakly excites
  14. L25:708 weakly excites
  15. L27:886 weakly suppresses
  16. L25:546 weakly excites
  17. L25:2912 weakly suppresses
  18. L26:1057 weakly excites
  19. L24:1657 weakly excites
  20. L25:916 weakly suppresses
  21. L24:1728 weakly excites
  22. L23:624 weakly excites
  23. L23:2728 weakly suppresses second digit of years like 2020
  24. L26:2946 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 -32.38 and -31.301 values between -26.99 and -25.911 values between -25.91 and -24.841 values between -23.76 and -22.681 values between -20.53 and -19.452 values between -18.37 and -17.301 values between -17.30 and -16.221 values between -16.22 and -15.141 values between -15.14 and -14.062 values between -12.99 and -11.914 values between -11.91 and -10.836 values between -10.83 and -9.765 values between -9.76 and -8.685 values between -8.68 and -7.6011 values between -7.60 and -6.5210 values between -6.52 and -5.4516 values between -5.45 and -4.3721 values between -4.37 and -3.2940 values between -3.29 and -2.22100 values between -2.22 and -1.141642 values between -1.14 and -0.061929 values between -0.06 and 1.0258 values between 1.02 and 2.0925 values between 2.09 and 3.1713 values between 3.17 and 4.255 values between 4.25 and 5.333 values between 5.33 and 6.402 values between 6.40 and 7.481 values between 9.63 and 10.71offnegativepositiveoften

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. Conan
  3. fibr
  4. itest
  5. .NoError
  6. fires
  7. BorderRadius
  8. ть
  9. antages
  10. atters

Suppresses

  1. 流
  2. school
  3. school
  4. flow
  5. School
  6. -school
  7. schools
  8. flows
  9. School
  10. SCHOOL