Neuron Explorer

Qwen3-0.6B

commas in non-English text

A neuron in layer 19 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 responds to commas, mostly in Russian and other non-English text, especially commas inside lists of words or before a clause that describes the words before it. When it was turned up, the model wrote long runs of 'which' and commas, so it seems tied to commas and relative clauses.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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

    j) право голосовать⏎2. Прочитайте текст Mayor-Council и заполните пропуски текста словами и выражениями, данными в

  2. #2 peak

    АБОТА ПО АНГЛИЙСКОМУ ЯЗЫКУ⏎1. Соотнесите значения слов и выражений, данных в левой колонке, с их английскими эквивал

  3. #3 peak

    ойте скобки, употребляя глаголы в Present, Past, Future Indef

  4. #4 peak

    ing Spiritualism and Materialism’, Ludwig Andreas Feuerbach wrote- “Der Mensch ist, was er ißt.” (Man is what he eats) and by this

  5. #5 peak

    ράφει τετράγωνον σχῆμα τῷ λόφῳ in Einklang zu bringen, ist es unumgänglich nothwendig dass der Ausdruck

  6. #6 peak

    pseudonymization,⏎i.e. processing it in a way that the personal data can no longer be mapped to a specific individual unless additional information is provided. Cases, where this can fail, including mapping of data from multiple sources. Customer activities recorded

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: the idea of a large group of people in a single place (such a place like the entire space) and the purpose of such a group in the context of the entire space—this is the beginning of the great philosophical and religious beginning.

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

    ического справочника Модальные глаголы и их эквиваленты и выполните следующее задание:⏎а) заполните пропуски<|endoftext|>Last Updated February

  2. #2 peak

    ЛИЙСКОМУ ЯЗЫКУ⏎1. Соотнесите значения слов и выражений, данных в левой колонке, с их английскими эквивалентами, данными в правой кол

  3. #3 peak

    голосовать⏎2. Прочитайте текст Mayor-Council и заполните пропуски текста словами и выражениями, данными в

  4. #4 peak

    ментарии при наличии в них нецензурной брани, оскорблений или сп

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. L18:247 excites
  2. L18:624 excites
  3. L17:18 excites unclear: mixed mid-sentence words
  4. L18:560 excites
  5. L18:804 suppresses
  6. L18:755 suppresses
  7. L18:823 excites
  8. L12:1006 suppresses
  9. L18:566 excites
  10. L10:112 suppresses
  11. L17:114 weakly excites
  12. L18:476 weakly excites
  13. L7:696 weakly excites
  14. L6:1656 weakly excites
  15. L18:477 weakly suppresses
  16. L17:19 weakly excites
  17. L17:398 weakly excites
  18. L17:895 weakly excites
  19. L18:382 weakly excites
  20. L17:714 weakly suppresses
  21. L17:1829 weakly suppresses
  22. L15:3035 weakly suppresses
  23. L18:1467 weakly suppresses
  24. L17:3 weakly excites

Feeds later layers

  1. L20:2891 strongly excites
  2. L22:368 strongly excites
  3. L20:811 strongly excites
  4. L20:1399 excites
  5. L26:236 excites
  6. L25:1221 excites
  7. L27:86 suppresses
  8. L22:940 excites
  9. L20:173 excites
  10. L26:22 suppresses the 'res' fragment inside words
  11. L20:882 suppresses
  12. L24:216 excites
  13. L21:632 excites
  14. L24:2151 excites
  15. L26:307 suppresses
  16. L25:328 excites
  17. L27:358 excites
  18. L21:809 excites
  19. L21:1070 excites
  20. L24:661 excites
  21. L24:304 excites
  22. L27:286 excites
  23. L25:438 excites
  24. L21:776 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 -1.32 and -1.254 values between -1.25 and -1.178 values between -1.17 and -1.096 values between -1.09 and -1.0110 values between -1.01 and -0.9410 values between -0.94 and -0.8613 values between -0.86 and -0.7816 values between -0.78 and -0.7018 values between -0.70 and -0.6318 values between -0.63 and -0.5530 values between -0.55 and -0.4745 values between -0.47 and -0.3968 values between -0.39 and -0.3196 values between -0.31 and -0.24138 values between -0.24 and -0.16257 values between -0.16 and -0.08437 values between -0.08 and -0.00535 values between -0.00 and 0.08555 values between 0.08 and 0.15459 values between 0.15 and 0.23357 values between 0.23 and 0.31274 values between 0.31 and 0.39201 values between 0.39 and 0.46129 values between 0.46 and 0.5482 values between 0.54 and 0.6260 values between 0.62 and 0.7036 values between 0.70 and 0.7716 values between 0.77 and 0.859 values between 0.85 and 0.938 values between 0.93 and 1.014 values between 1.01 and 1.096 values between 1.09 and 1.161 values between 1.32 and 1.401 values between 1.71 and 1.78offnegativepositiveoften

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. ,function
  3. which
  4. ，
  5. ,tr
  6. ，在
  7. ，并
  8. Which
  9. ,'#
  10. который

Suppresses

  1. filer
  2. tac
  3. rame
  4. zens
  5. intosh
  6. まず
  7. mez
  8. odore
  9. nen
  10. stell