Neuron Explorer

Qwen3-0.6B

the gap before a decade or century

A neuron in layer 4 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 small space between 'late' and a year or century that follows, as in 'the late 1930s' or 'the late 18th century'. It is a narrow, predictable trigger. Turning it up did not produce writing that clearly related to dates.

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

    the late 1930s skiers have been trying to open up the San G

  2. #2 peak

    the late 18th century, with the mechanisation of the textile industry. Tasks previously

  3. #3 peak

    allando. Alphabet design became a subject of study, and several technical treatises were published on writing styles.⏎By the late 16th cent., with the secure establishment of the printing press, the art

  4. #4 peak

    Staircase-Escalante National Monument) to find out more about paleontology in that area.<|endoftext|>The principal systems of units used in scientific work, late 19th – early 20th centuries. In the United States,

  5. #5 peak

    imaging from space.⏎Rapid technological developments soon led to researchers receiving close-up views of the Moon and the inner planets. Within a few decades, by the late 1970s, planetary researchers were receiving thousands of colour images from Mars

  6. #6 peak

    ballad in four stanzas, as well as the melody for the ballad, appears in one of the key sources of English popular song of the late 17th and early 18th century called ‘Wit and M

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: it's the only thing that matters. This is a common approach of using the term "apple" to represent the solution. Now, I will have to be the most important. In the context of the term, let's say we are

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

    system, and multithreading.⏎Also included is coverage of some of Java's most powerful features,⏎such as generics, lambda expressions, and the very important⏎modules feature. A number of recent

  2. #2 peak

    foundational features (such as data types, operators,⏎control statements, classes, and interfaces) to its advanced topics⏎(such as multithreading, annotations, generics, lambda expressions,⏎and modules). Of course, recent

  3. #3 peak

    Act implementation; (2) assess changes in the proportion of schools with excellent water access and barriers and facilitators to More<|endoftext|>Minnesota is a disclose in the Upper Midwest,

  4. #4 peak

    becoming constant. It may also become more severe when coughing, sneezing, or moving around. The pain may eventually radiate to the lower right side of the abdomen, where the appendix is located.⏎Loss of Appetite⏎One sign

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. L1:1385 suppresses
  2. L0:1224 suppresses
  3. L0:329 suppresses the digit 1 in years
  4. L3:50 weakly excites
  5. L0:178 weakly suppresses
  6. L3:115 weakly suppresses
  7. L3:59 weakly excites
  8. L3:122 weakly excites
  9. L1:1698 weakly suppresses
  10. L0:1839 weakly suppresses
  11. L2:1657 weakly excites
  12. L3:2998 weakly suppresses
  13. L3:464 weakly excites
  14. L0:1301 weakly suppresses
  15. L3:2042 weakly excites
  16. L0:2751 weakly suppresses
  17. L3:1484 weakly excites
  18. L3:1252 weakly excites
  19. L2:1494 weakly suppresses
  20. L1:1142 weakly suppresses
  21. L2:75 weakly suppresses
  22. L1:1999 weakly excites
  23. L0:407 weakly suppresses
  24. L1:2071 weakly suppresses

Feeds later layers

  1. L5:659 suppresses
  2. L5:788 excites
  3. L8:1478 suppresses
  4. L5:1656 excites 'of' after end, top or bottom
  5. L5:3024 suppresses
  6. L6:890 excites
  7. L7:558 weakly suppresses
  8. L12:2853 weakly excites
  9. L14:2658 weakly suppresses
  10. L7:2088 weakly excites
  11. L7:1941 weakly excites
  12. L12:272 weakly excites
  13. L5:1342 weakly suppresses
  14. L19:2872 weakly excites
  15. L25:1985 weakly suppresses
  16. L22:546 weakly excites
  17. L6:1167 weakly excites
  18. L18:2756 weakly suppresses
  19. L20:299 weakly suppresses
  20. L25:539 weakly suppresses
  21. L8:485 weakly suppresses
  22. L11:859 weakly excites
  23. L22:1248 weakly excites
  24. L22:2027 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.

28 values between -0.64 and -0.38629 values between -0.38 and -0.122949 values between -0.12 and 0.14273 values between 0.14 and 0.4010 values between 0.40 and 0.665 values between 0.66 and 0.922 values between 0.92 and 1.181 values between 1.18 and 1.442 values between 1.44 and 1.701 values between 1.96 and 2.221 values between 2.48 and 2.741 values between 2.74 and 3.002 values between 3.00 and 3.261 values between 3.78 and 4.041 values between 4.55 and 4.811 values between 5.33 and 5.591 values between 9.48 and 9.74offnegativepositiveoften

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. erin
  2. ่าน
  3. 령
  4. olson
  5. erdem
  6. -domain
  7. buch
  8. eri
  9. rencontre
  10. abeth

Suppresses

  1. minor
  2. T
  3. cha
  4. �回
  5. minor
  6. ños
  7. 每次
  8. GRA
  9. �试
  10. *