Neuron Explorer

Qwen3-0.6B

hyphen inside compound words

A neuron in layer 10 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 hyphens that join or split a word, such as 'breast-feeding' or 'veg-' at a line break. It is strongly negative on digits inside years and copyright dates. Turning it up did not produce a clear change in the writing.

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

    Diets lower blood pressure for Children . 12 An impressive number of studies, dating back to the early 1920s, show that veg-<|endoftext|>If you take a walk through the Salt Lake City neighborhoods and hills around our

  2. #2 peak

    in chemically reacting powder mixtures⏎- pressure processes which are descriptive of "ballotechnic" reactions, i.e. shock- induced reactions in heterogeneous ma- terial mixtures.⏎- ISOTOPIC SEPARATION AND ENR

  3. #3 peak

    breast milk.⏎Henderson, a mother of two, says she wasn't too conscious of what she ate while breast-feeding, but has noticed a difference in the preferences of children who breast-feed.⏎

  4. #4 peak

    or ask your own question.⏎Posted 31 Aug 2009 • 1 answer⏎Posted 20 Jan 2010 • 1 answer⏎Posted 27 Jun 2010 • 1

  5. #5 peak

    DS. AROMATIC COMPOUNDS ARE THOSE ORGANIC COMPOUNDS WHICH HAVE BENZENE RING (CYCLOHEXA-1,3,5-TRIENE). AROMATIC COMPOUN

  6. #6 peak

    Sep 2012 • 1 answer⏎Posted 3 Oct 2014 • 1 answer<|endoftext|>Choosing Whole Foods for a Healthier You⏎Written by Eric De

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 to do it. I have to do it, and I have to do it. I have200000000000000000000000000

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

    : Barnave's Introduction to the French Revolution and Other Writings (1971).⏎The Columbia Electronic Encyclopedia, 6th ed. Copyright © 2012, Columbia University Press. All rights reserved.⏎See more Encyclopedia

  2. #2 peak

    the Nobel Prize in 1903 for her work science and chemistry. She discovered two new elements including rad

  3. #3 peak

    Enchanted Learning website for:|⏎Copyright ©2002-2016 EnchantedLearning.com ------<|endoftext|>Next analogy uses the fresh etymology

  4. #4 peak

    into the pockets and mouths of students, drop-outs, and "free thinkers".⏎And soon it drew the attention of the authorities. By the summer of 1966, the home secretary had moved to ban it

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. L4:53 excites blank slot before a number
  2. L9:349 suppresses the spot before a year
  3. L8:111 suppresses the space before a year in dates
  4. L9:13 excites
  5. L9:2283 suppresses
  6. L8:457 excites year digits in copyright and dates
  7. L6:371 suppresses
  8. L6:2086 weakly excites
  9. L9:974 weakly suppresses
  10. L8:421 weakly excites
  11. L4:987 weakly suppresses
  12. L7:229 weakly excites first digit of a split number
  13. L3:3 weakly suppresses
  14. L7:6 weakly suppresses
  15. L0:602 weakly excites words right before a number
  16. L3:670 weakly suppresses
  17. L1:2733 weakly excites
  18. L8:2591 weakly excites
  19. L8:2569 weakly excites
  20. L6:4 weakly suppresses
  21. L6:283 weakly excites
  22. L4:318 weakly excites the digit 9 in years 1900s
  23. L8:1165 weakly excites
  24. L8:232 weakly suppresses

Feeds later layers

  1. L16:256 suppresses number right after 'age of' or 'ages'
  2. L14:140 suppresses
  3. L16:396 suppresses
  4. L17:222 suppresses
  5. L17:170 suppresses
  6. L17:15 excites
  7. L17:285 suppresses
  8. L18:377 suppresses number right after 'about' or 'less than'
  9. L14:102 excites
  10. L14:616 suppresses
  11. L20:4 suppresses
  12. L17:126 suppresses
  13. L23:12 suppresses
  14. L16:723 excites
  15. L21:70 suppresses
  16. L14:926 suppresses
  17. L18:991 suppresses
  18. L25:182 suppresses
  19. L18:91 suppresses
  20. L18:1534 suppresses space between month and day number
  21. L16:1044 suppresses
  22. L21:38 suppresses
  23. L18:2578 suppresses
  24. L11:15 suppresses decade and century numbers like 1950s

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 -9.41 and -9.151 values between -8.62 and -8.352 values between -8.35 and -8.091 values between -8.09 and -7.821 values between -7.56 and -7.292 values between -7.29 and -7.021 values between -7.02 and -6.761 values between -6.76 and -6.493 values between -6.23 and -5.963 values between -5.96 and -5.704 values between -5.70 and -5.433 values between -5.43 and -5.174 values between -5.17 and -4.906 values between -4.90 and -4.636 values between -4.63 and -4.374 values between -4.37 and -4.102 values between -4.10 and -3.842 values between -3.84 and -3.571 values between -3.57 and -3.317 values between -3.31 and -3.045 values between -3.04 and -2.785 values between -2.78 and -2.517 values between -2.51 and -2.243 values between -2.24 and -1.983 values between -1.98 and -1.714 values between -1.71 and -1.459 values between -1.45 and -1.1814 values between -1.18 and -0.927 values between -0.92 and -0.6512 values between -0.65 and -0.3972 values between -0.39 and -0.121792 values between -0.12 and 0.151710 values between 0.15 and 0.41199 values between 0.41 and 0.689 values between 0.68 and 0.942 values between 0.94 and 1.21offnegativepositiveoften

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

  2. }
  3. 送出
  4. Ta
  5. gan
  6. soft
  7. Le
  8. 藉
  9. ta
  10. Te

Suppresses

  1. osoph
  2. ónico
  3. agnost
  4. 的人来说
  5. osopher
  6. ertime
  7. ewis
  8. <=$
  9. 万户
  10. ,callback