Neuron Explorer

Qwen3-0.6B

period inside country abbreviations

A neuron in layer 7 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 full stop inside abbreviations such as 'U.S.', 'U.K.' and 'U.S.C.', where the dot is part of a name and not the end of a sentence. Turning it up did not produce any clear theme in the model's 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

    cited statute is Section 1185 of Title 23 of the Vermont Statutes Annotated. And the federal citation 42 U.S.C. § 1395 tells us that this cited federal statute can be found

  2. #2 peak

    official seal of the CIA, however, may NOT be copied without permission as required by the CIA Act of 1949 (50 U.S.C. section 403m). Misuse of the official seal of the CIA

  3. #3 peak

    Wednesday.⏎The legal case comes after last week when climate advisors in the U.K. warned that the younger generation will be the ones paying the price of climate change.

  4. #4 peak

    sunscreen are more likely to experience painful, damaging sunburns, a new study shows.⏎The study, the first to look at how people in the U.S. shield themselves from the sun and how well those different strategies work, analyzed information on

  5. #5 peak

    their students. Texas, Michigan, Idaho, Illinois and Virginia also set a low bar.⏎Standards still declined in rigor in 26 states and D.C. between 2009 and 2011, while 2

  6. #6 peak

    to 70 mph at Austin, TX, and gusts to 75 mph at Tulsa, OK.⏎Many cities in the central and eastern U.S. reported record high temperatures for the date. Highs of 100°

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: problems. The 2010 problems. Unselected. Triangles. Triangles. Numbers. Triangles. Numbers. L. V. Tri. Numbers. L. V. Tri. Numbers. L. V. Numbers.

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

    reproduced in⏎Rice: Amber the Golden Gem of the Ages, the Polish Academy of Sciences:⏎Amber in Nature, a German⏎publication: Lausitzer Bernstein, and Grimaldi:⏎Amber Window to the Past.⏎Mark Meyer

  2. #2 peak

    ldesammlungen, Munich, Germany|⏎|Credit||Courtesy of Bildarchiv Preussischer Kulturbesitz/Art Resource, NY|⏎|Larry SilverProfessor of

  3. #3 peak

    cause of one of the botanical mysteries of the Burren as the Dense Flowered Orchid is native of the Mediterranean yet here growing alongside the (Alpine) Gentian and often with Mountain

  4. #4 peak

    in large quantities. Ceylon (Sri Lankan) cinnamon has lower levels of coumarin.⏎A study by researcher J.H.

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. L5:1747 suppresses
  2. L5:125 excites "of" after all three, both, ten
  3. L6:1718 suppresses
  4. L6:3044 suppresses
  5. L4:2743 excites
  6. L4:2335 excites
  7. L6:972 excites "well" after "how"
  8. L6:479 excites
  9. L6:1789 suppresses
  10. L6:3016 suppresses
  11. L5:1372 excites
  12. L6:199 excites
  13. L4:212 excites
  14. L5:1788 suppresses
  15. L6:1540 excites
  16. L6:181 weakly excites periods in abbreviations and the 's in names
  17. L4:285 weakly excites
  18. L2:1126 weakly suppresses
  19. L1:818 weakly suppresses
  20. L4:951 weakly suppresses
  21. L4:19 weakly suppresses
  22. L1:682 weakly excites
  23. L1:761 weakly excites
  24. L4:2473 weakly excites

Feeds later layers

  1. L10:167 strongly suppresses
  2. L8:252 excites 'of' after 'each' or 'one' or 'all'
  3. L17:985 suppresses
  4. L15:557 excites
  5. L11:810 excites
  6. L18:602 suppresses
  7. L13:975 suppresses
  8. L19:663 suppresses
  9. L21:274 excites
  10. L13:527 suppresses
  11. L20:337 excites
  12. L13:641 suppresses
  13. L17:1027 suppresses
  14. L14:2271 excites
  15. L9:240 suppresses
  16. L8:2755 suppresses
  17. L9:152 suppresses
  18. L21:68 excites
  19. L18:277 excites
  20. L12:372 suppresses
  21. L8:393 suppresses
  22. L16:2520 suppresses
  23. L9:434 suppresses comma after an adjective in a list
  24. L8:2095 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 -4.22 and -4.081 values between -2.06 and -1.912 values between -1.19 and -1.053 values between -1.05 and -0.912 values between -0.91 and -0.765 values between -0.76 and -0.625 values between -0.62 and -0.4737 values between -0.47 and -0.33236 values between -0.33 and -0.181076 values between -0.18 and -0.041863 values between -0.04 and 0.11527 values between 0.11 and 0.2599 values between 0.25 and 0.3928 values between 0.39 and 0.548 values between 0.54 and 0.687 values between 0.68 and 0.831 values between 0.83 and 0.972 values between 0.97 and 1.121 values between 1.12 and 1.261 values between 1.26 and 1.403 values between 1.40 and 1.55offnegativepositiveoften

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. lessness
  2. less
  3. weeney
  4. 峨
  5. .blit
  6. ฆ
  7. erne
  8. 拜师学艺
  9. -less
  10. 蚓

Suppresses

  1. 时候
  2. ことが
  3. älle
  4. transformations
  5. appen
  6. /**<
  7. ことに
  8. recurring
  9. oor
  10. validators