Neuron Explorer

Qwen3-0.6B

'Ch' starting a proper name

A neuron in layer 27 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 'Ch' at the start of a word that is a place or name, such as Chomolungma or Chittenden, when the word has been split into pieces. Turning it up made the model output blank lines, and turning it down made it wander into unrelated text.

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

    Government of Nepal gave the mountain the official Nepali name. In sanskrit Sagarmatha means "mother of the universe”. The Tibetan name for Mount Everest is Chomolungma or

  2. #2 peak

    are smarter, more efficient, more earth-friendly, and which incorporate a variety of beautiful new forms not seen before in inhabited structures.⏎Our final image is of the Chesa Futura in St. Moritz, Switzerland, a futuristic ski chalet

  3. #3 peak

    25th of October, that the British cavalry were drawn up under Lord Lucan at the western end of this narrow valley directly under the steep heights of the Chersonese. On the summit, at the very edge of the heights, Lord

  4. #4 peak

    celebrated English writer.⏎Capital, Middlebury. Population, 26,549.⏎is the chief town. This county is bounded on the north by Chittenden⏎County, east by Washington and Orange Counties, and a part of

  5. #5 peak

    in them with ease.⏎There are currently 11 national parks in the Thar Desert, the largest of which are the Rann of Kutch and the Nara Desert Wildlife Sanctuaries. The Rann of Kutch is the only

  6. #6 peak

    49, Taiwan (e-mail: firstname.lastname@example.org).⏎Cite:Ming-Ju Lee, Ming-Gin Lee, Yishuo Huang, and Chia-Liang Chiang, "Purification Study of Pervious Concrete

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: there are plenty of sneraries snearing sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers sneers

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

    Is your person [name]?”.⏎Players are quick to understand that by asking questions that divide the candidates into two groups, such as “is your person female?”, you can reduce the candidate pool more quickly.⏎If every question halved the

  2. #2 peak

    find 386 pages of resources to learn more about Caddo history and culture, https://scholarworks.sfasu.edu/cgi/viewcontent.cgi?article=4012&context=ita⏎Visit the THC’s Learning

  3. #3 peak

    and research. Generally, these conversations can be split into two types: “Why can’t I get genetic testing?” and “What’s the point of genetic testing? It’s‘not actionable.’”⏎The American Academy of Pediatrics, along with other

  4. #4 peak

    “What is the meaning of Menstrual Hygiene Day?”, “How did it begin?”, “Why is Menstrual Hygiene Day Celebrated?”, and “How can you commemorate Menstrual Hygiene Day?”⏎Men

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. L21:1894 strongly excites
  2. L26:193 strongly suppresses
  3. L25:118 excites
  4. L26:130 excites ellipsis cut-off in quotes
  5. L2:202 suppresses "of" after "many" or "which"
  6. L23:474 excites
  7. L20:863 excites
  8. L16:41 excites
  9. L25:483 excites
  10. L16:47 excites
  11. L2:1753 excites
  12. L19:89 suppresses
  13. L0:117 excites
  14. L25:36 suppresses
  15. L14:51 excites colon that introduces a statement
  16. L15:123 suppresses
  17. L18:48 suppresses
  18. L26:725 excites
  19. L17:118 excites comma after a clause ending
  20. L21:173 excites
  21. L20:61 suppresses
  22. L22:271 suppresses
  23. L13:10 excites
  24. L20:8 suppresses

Feeds later layers

None: this is the last layer.

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 -59.63 and -57.862 values between -41.95 and -40.191 values between -31.35 and -29.581 values between -15.45 and -13.681 values between -13.68 and -11.911 values between -11.91 and -10.141 values between -10.14 and -8.383 values between -8.38 and -6.612 values between -6.61 and -4.847 values between -4.84 and -3.0852 values between -3.08 and -1.313689 values between -1.31 and 0.46143 values between 0.46 and 2.232 values between 2.23 and 3.991 values between 3.99 and 5.761 values between 9.30 and 11.06offnegativepositiveoften

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. ⏎⏎
  3. </
  4. ”
  5. ”⏎
  6. )
  7. ⏎ ⏎
  8. ⏎→→→⏎
  9. ⏎→→→→⏎
  10. )”

Suppresses

  1. .go
  2. raries
  3. tein
  4. 清醒
  5. criptive
  6. iciones
  7. eric
  8. romatic
  9. confisc
  10. /domain