Neuron Explorer

Qwen3-0.6B

the 's' after a hyphen

A neuron in layer 27 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires strongly on the letter 's' that comes right after a hyphen in compound words, such as 'fat-soluble', 'sugar-sweetened' or 'alpha-synuclein'. Turned up, the model repeated words like 'our' and 'representative'. Turned down, it repeated 'clearest' and 'likely'.

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

    can synthesize the micronutrient when you’re exposed to unfiltered sunlight. Inadequate levels of vitamin D can have serious health consequences.⏎As a fat-soluble vitamin -- others are A, E and K -- vitamin D can be

  2. #2 peak

    Parkinson's and obstructive sleep apnea.⏎- In an animal model of Parkinson's, researchers found that a high-sugar diet did not induce neurodegeneration and protected from induced neurodegeneration caused

  3. #3 peak

    of the brain responsible for much of our memory and learning functions. Our hippocampal volume naturally decreases with age, but this study suggests that regular consumption of sugar-sweetened beverages may accelerate this process.⏎Diet soda can age your brain too

  4. #4 peak

    ⏎- you have low levels of sodium in your body – this can happen if you're dehydrated or if you're on a low-sodium (low-salt) diet⏎- you have Addison's disease, a rare disorder of the

  5. #5 peak

    early stages of Parkinson's-related neurodegeneration. These findings are "striking," because alpha-synuclein has long been thought to be the more significant driver of Parkinson's

  6. #6 peak

    formal imposition of Christian authority on an Islamic society.⏎Valencia's Muslims<|endoftext|>Trauma remains a leading cause of death and disability in adults in spite of advances in resuscitation, surgical management, and critical care . Between 25 to

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: Maybe all possible actionsctrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrinectrine

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

    for adult) in vein below the tourniquet⏎- Dorsal common digital vein (dorsum)⏎- Lateral digital vein (near dewclaws)⏎How does infection spread from the coffin joint, and what does it cause

  2. #2 peak

    steel, wire and concrete and ornamented the three main spires and their 14 surrounding sculptural elements with colorful bits of broken glass, pottery and seashells. His work is an example of non-traditional vernacular architecture and

  3. #3 peak

    , Caroline Benedicks⏎. They fell in love, married, and eventually established themselves in a permanent year-round residence, Brucebo⏎, on the western seashore of Gotland Island in Sweden. From their artistic union two significant institutions in

  4. #4 peak

    , from the movie “The Little Mermaid”<|endoftext|>Jan Brett's Home Page⏎The Seashell Rounding Activity⏎Rounding makes numbers easier to⏎work with in your

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. L26:1564 excites
  2. L22:164 suppresses
  3. L25:206 excites
  4. L24:696 excites
  5. L24:105 suppresses
  6. L25:578 suppresses
  7. L18:569 excites
  8. L21:746 suppresses
  9. L21:545 excites
  10. L23:206 suppresses
  11. L1:2677 suppresses
  12. L8:16 weakly suppresses
  13. L19:660 weakly suppresses
  14. L18:1115 weakly suppresses
  15. L20:959 weakly excites
  16. L16:248 weakly excites
  17. L26:495 weakly suppresses
  18. L18:1116 weakly excites
  19. L14:2974 weakly suppresses
  20. L19:2076 weakly excites
  21. L21:1326 weakly excites capital 'S' starting a name
  22. L18:871 weakly excites
  23. L8:944 weakly suppresses
  24. L17:1717 weakly excites

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 -49.22 and -46.631 values between -38.88 and -36.301 values between -28.54 and -25.961 values between -25.96 and -23.381 values between -23.38 and -20.795 values between -20.79 and -18.211 values between -18.21 and -15.622 values between -15.62 and -13.044 values between -13.04 and -10.456 values between -10.45 and -7.878 values between -7.87 and -5.2810 values between -5.28 and -2.70403 values between -2.70 and -0.123433 values between -0.12 and 2.4710 values between 2.47 and 5.055 values between 5.05 and 7.647 values between 7.64 and 10.222 values between 10.22 and 12.811 values between 12.81 and 15.392 values between 17.98 and 20.562 values between 20.56 and 23.141 values between 28.31 and 30.901 values between 51.57 and 54.16offnegativepositiveoften

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. or
  2. cập
  3. 绁
  4. Our
  5. our
  6. ;*/⏎
  7. века
  8. tain
  9. orical
  10. 已久的

Suppresses

  1. _proba
  2. obs
  3. ctrine
  4. nist
  5. Bloc
  6. ROP
  7. še
  8. likely
  9. 全额
  10. xDD