Neuron Explorer

Qwen3-0.6B

unclear: mid-word fragments, often 'char'

A neuron in layer 24 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

It fires on a fragment in the middle of a split word, such as the piece of 'ugali' or 'unsigned', and it is strongest in the middle of a word rather than at the start. Turning it up made the model repeat 'char' and similar code-type words, while turning it down pushed the text toward unrelated phrases.

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

    food such as ugali. Common traditional vegetables include⏎baobab, cowpeas, am

  2. #2 peak

    (RBM), deep belief networks (DBN) and deep convolutional neural networks (CNN). In recent years, CNN-based methods have gained popularity<|endoftext|>typedef unsigned int size_t;⏎This type is used to hold size definitions. Since sizes are

  3. #3 peak

    have occurred in areas with an abundance of young rabbits. My sightings have been in the morning hours or at dusk. When my wife and I were in the High Uintahs, we witnessed a Canadian Lynx. It was in an area where

  4. #4 peak

    ate character in block of memory⏎Searches within the first num bytes of the block of memory pointed by ptr for the first occurrence of value (interpreted as an unsigned char), and returns a pointer to it.⏎Both value and each of the bytes

  5. #5 peak

    the labor force and future careers. Please be sure to visit www.LL-inc.org for activities to facilitate career-conversation with your young leader.<|endoftext|>How does Utterson perceive the relationship between Jekyll and Hyde for most of the novel? Is

  6. #6 peak

    ’re concerned about a child’s vaping habits. In case you have any concerns about exactly where and the way to utilize น้ำยาบุหรี่ไฟฟ้าราคาส่ง ยก ลัง, you can e

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: we can take 20000000000000000000000000000000000000000000

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

    , which looms on the horizon north of the major city of Yogyakarta, is on the island of Java.⏎President Susilo Bambang Yudhoyono has called the volcanic eruptions a crisis situation. He and several

  2. #2 peak

    of Indonesia's most active volcanoes and lies in one of the world's most densely populated areas. Many people raise crops and livestock in its shadow.⏎Yudhoyono has announced that residents will receive compensation for livelihoods and animals lost to

  3. #3 peak

    our democracy for better results throughout the history.⏎Influence by different rulers, the city has been home to different design masterpiece such as Indraparastha, Daulatabad, Old Delhi and many more. Sur

  4. #4 peak

    prevents damage from overloading. Then there is a bridge rectifier made of four 1N4007 PN junction diodes and a filter capacitor of 2.2uF/450V.⏎This<|endoftext|>Warning: Use

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. L23:1647 excites
  2. L22:563 suppresses
  3. L21:2222 suppresses
  4. L20:936 suppresses
  5. L21:448 weakly excites
  6. L17:343 weakly suppresses
  7. L4:368 weakly excites
  8. L9:214 weakly excites
  9. L6:1313 weakly excites
  10. L20:431 weakly suppresses
  11. L4:2089 weakly suppresses
  12. L20:2148 weakly excites
  13. L20:669 weakly suppresses
  14. L5:173 weakly excites
  15. L11:1402 weakly excites
  16. L1:2015 weakly suppresses
  17. L1:419 weakly suppresses
  18. L19:542 weakly suppresses
  19. L18:3020 weakly excites
  20. L19:2093 weakly excites
  21. L11:840 weakly suppresses
  22. L22:1136 weakly excites
  23. L23:728 weakly excites
  24. L9:2068 weakly excites

Feeds later layers

  1. L27:1623 weakly excites
  2. L25:2905 weakly suppresses
  3. L26:541 weakly excites
  4. L27:1049 weakly excites
  5. L25:260 weakly suppresses
  6. L25:2142 weakly excites
  7. L26:1463 weakly suppresses
  8. L25:1606 weakly excites
  9. L27:1073 weakly excites
  10. L25:1380 weakly suppresses
  11. L27:2872 weakly excites
  12. L25:398 weakly excites
  13. L25:1994 weakly excites
  14. L25:2029 weakly excites
  15. L25:1261 weakly excites
  16. L25:511 weakly excites
  17. L25:1349 weakly excites
  18. L25:2847 weakly excites
  19. L27:2659 weakly excites
  20. L26:1867 weakly suppresses
  21. L25:2491 weakly excites
  22. L26:2445 weakly suppresses
  23. L25:2522 weakly suppresses
  24. L25:1118 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.

2 values between -16.42 and -15.392 values between -6.06 and -5.021 values between -5.02 and -3.982 values between -3.98 and -2.957 values between -2.95 and -1.9132 values between -1.91 and -0.873164 values between -0.87 and 0.17662 values between 0.17 and 1.2016 values between 1.20 and 2.244 values between 2.24 and 3.286 values between 3.28 and 4.312 values between 4.31 and 5.351 values between 5.35 and 6.392 values between 6.39 and 7.421 values between 7.42 and 8.462 values between 13.64 and 14.681 values between 17.79 and 18.831 values between 24.01 and 25.05offnegativepositiveoften

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. char
  2. char
  3. →char
  4. (char
  5. Char
  6. Char
  7. byte
  8. [char
  9. -char
  10. Character

Suppresses

  1. перв
  2. verity
  3. uspended
  4. oins
  5. 社区
  6. nous
  7. YGON
  8. 恋
  9. ła
  10. Community