Neuron Explorer

Qwen3-0.6B

end-of-document token boundary

A neuron in layer 17 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 most strongly on the marker that signals the end of one web document and the start of another. It goes quiet, or negative, on dashes and colons inside citation or metadata lines. Turning it up did not give a clear shared theme in the text it wrote.

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

    liu.<|endoftext|>Bond Price Calculator⏎With this bond price calculator, we aim to help you calculate

  2. #2 peak

    stricting blood vessels?<|endoftext|>There is growing recognition that a sedentary lifestyle is being driven, at least

  3. #3 peak

    ,0<|endoftext|>Guidelines for Restoring Cultural Landscapes⏎Identifying, retaining and preserving

  4. #4 peak

    , 2012<|endoftext|>Some weather conditions are simply not fun to ride in. As cyclists, windy days

  5. #5 peak

    Influence Insider?<|endoftext|>As the world scrambles to find antiviral treatments and vaccines for the novel

  6. #6 peak

    to review it. We're adding the slope at a single point to the slope over a big area, which is something I can't quite work out visually.<|endoftext|>The world produces more than enough food to feed everyone, yet each night more than

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 have a 300-100-100-100-100-100-100-1-10-1-1-1-1-1-1

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

    backgrounds.⏎|Original language||English (US)|⏎|Number of pages||12|⏎|Journal||Classical and Quantum Gravity|⏎|State||Published - Jul 7 2003|⏎ASJC Scopus subject areas⏎

  2. #2 peak

    ENNEDY, Chi Kin John LEE|⏎|Place of Publication||Abingdon, Oxon|⏎|Publication status||E-pub ahead of print - Aug 2020|<|endoftext|>10. Rebinding an Old Book

  3. #3 peak

    and A[1:0] for ARM instruction fetches (nOPC=0 and MAS[1:0]=10).⏎Article last edited on: 2008-09-09 15:4

  4. #4 peak

    Jiangsu 211171, China⏎*Authors to whom correspondence should be addressed.⏎Received: 2020-8-4 / Accepted: 2020-9-8 / Published: 2

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. L8:4 suppresses
  2. L8:13 excites
  3. L8:7 suppresses
  4. L11:2 excites unclear: blank before a year or number
  5. L8:16 excites
  6. L6:13 suppresses
  7. L8:10 suppresses
  8. L6:12 suppresses
  9. L8:3 suppresses
  10. L9:5 suppresses
  11. L3:464 suppresses
  12. L7:64 excites
  13. L7:191 suppresses
  14. L6:7 excites
  15. L5:50 suppresses
  16. L8:14 suppresses
  17. L7:102 excites
  18. L6:18 suppresses
  19. L8:12 suppresses
  20. L10:46 suppresses
  21. L13:129 suppresses
  22. L5:158 excites
  23. L15:5 suppresses
  24. L7:122 suppresses

Feeds later layers

  1. L27:725 weakly suppresses
  2. L27:84 weakly excites
  3. L27:1665 weakly excites
  4. L27:1974 weakly excites
  5. L27:2211 weakly excites
  6. L27:314 weakly excites
  7. L27:1995 weakly excites
  8. L27:2116 weakly excites unclear: first and boundary tokens
  9. L27:82 weakly excites
  10. L27:2326 weakly excites
  11. L27:107 weakly excites
  12. L19:423 weakly excites words that set up a following clause
  13. L27:194 weakly excites
  14. L19:669 weakly excites
  15. L27:1371 weakly excites
  16. L27:1 weakly suppresses
  17. L27:47 weakly excites
  18. L27:823 weakly suppresses
  19. L27:37 weakly excites
  20. L27:1345 weakly suppresses
  21. L27:301 weakly excites
  22. L27:1466 weakly excites
  23. L27:2110 weakly suppresses
  24. L27:801 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.

1 values between -12.42 and -11.781 values between -11.78 and -11.154 values between -11.15 and -10.512 values between -10.51 and -9.877 values between -9.87 and -9.239 values between -9.23 and -8.5915 values between -8.59 and -7.9522 values between -7.95 and -7.3234 values between -7.32 and -6.6893 values between -6.68 and -6.04109 values between -6.04 and -5.40180 values between -5.40 and -4.76299 values between -4.76 and -4.12355 values between -4.12 and -3.49459 values between -3.49 and -2.85529 values between -2.85 and -2.21563 values between -2.21 and -1.57541 values between -1.57 and -0.93371 values between -0.93 and -0.30191 values between -0.30 and 0.3458 values between 0.34 and 0.9830 values between 0.98 and 1.6216 values between 1.62 and 2.265 values between 2.26 and 2.903 values between 2.90 and 3.542 values between 3.54 and 4.172 values between 4.17 and 4.811 values between 4.81 and 5.451 values between 5.45 and 6.091 values between 6.09 and 6.731 values between 7.37 and 8.001 values between 8.00 and 8.641 values between 9.28 and 9.921 values between 12.47 and 13.11offnegativepositiveoften

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. ilian
  2. otional
  3. uggage
  4. usal
  5. ikers
  6. Wants
  7. ughters
  8. ailer
  9. Board
  10. htags

Suppresses

  1. 举
  2. правило
  3. именно
  4. recourse
  5. 打击
  6. —it
  7. assortment
  8. <source
  9. plaintext
  10. нос