Neuron Explorer

Qwen3-0.6B

unclear: fragments of split words

A neuron in layer 27 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 random pieces of longer words that the model has split apart, such as the start of 'splints' or 'sorrel', with no clear shared topic. Turning it down made the model repeat the word 'vertex', and turning it up gave rambling text.

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

    a religion is necessarily, from the scholastic point of view, the proper person to speak for the religion. They go on to characterize as a "tightrope between academic and religious worlds of value" the situation of a scholar who has

  2. #2 peak

    . The other four pairs are extrapulmonary prolongations of secondary bronchi: (1) cervical, lying along the esophagus; (2) clavicular, frequently merging into a single interclavicular sac; (3

  3. #3 peak

    (family Polygonaceae). In California, sheep sorrel occurs in a freshwater marsh community and is common in annual grasslands, mont

  4. #4 peak

    humerus. The posterior fat pad is not seen because it lies in the olecranon fossa of the distal humerus, depressed superficially by the triceps tendon and anconeus muscle. Traumatic effusions will lift the posterior

  5. #5 peak

    tibia) hurts when you press on it firmly with your finger.⏎The causes of shins splints include problems with body mechanics, a sudden or excessive increase in the amount or

  6. #6 peak

    important to figure out the level of severity (stress injury or stress fracture) in order to determine the right treatment. If you’ve had shin splints before, you can reduce the risk of a recurrence by getting the right shoe

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: vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex vertex

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

    security on farms⏎Biosecurity is about reducing the risk of disease transmission. We know that bTB can be transmitted by direct contact (e.g. nose-to-nose) or indirect contact (e.g. via contaminated pasture) but simple exclusion

  2. #2 peak

    Benin, Colombia, Djibouti, Greece, Moldova, Mongolia, and Senegal.⏎Most

  3. #3 peak

    or under drought conditions. The tubers are most susceptible, but stems and roots may occasionally be infected. No above-ground symptoms are apparent on infected potato plants and scab does not adversely affect potato tubers in storage.⏎The bacterium survives in

  4. #4 peak

    Pat skin dry and immediately apply a layer of moisturizer while skin is still slightly damp. Choose one that contains ingredients known to be effective for dry skin such as hyaluronic acid, glycerin, alpha-hydroxy acids or<|endoftext|>

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. L23:2598 excites
  2. L25:1660 suppresses
  3. L22:2376 excites
  4. L6:13 excites
  5. L26:1468 suppresses
  6. L22:1836 weakly suppresses
  7. L24:34 weakly excites
  8. L23:319 weakly suppresses
  9. L22:2479 weakly suppresses
  10. L7:110 weakly suppresses
  11. L26:1564 weakly excites
  12. L12:2069 weakly excites 'War' after 'World'
  13. L26:453 weakly suppresses
  14. L24:385 weakly excites
  15. L0:2322 weakly suppresses
  16. L22:2487 weakly excites
  17. L23:206 weakly suppresses
  18. L20:1515 weakly excites
  19. L25:319 weakly suppresses
  20. L8:393 weakly suppresses
  21. L8:14 weakly excites
  22. L9:2225 weakly suppresses
  23. L6:16 weakly excites
  24. L2:75 weakly 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 -24.17 and -22.152 values between -12.03 and -10.013 values between -10.01 and -7.985 values between -7.98 and -5.9611 values between -5.96 and -3.9315 values between -3.93 and -1.911854 values between -1.91 and 0.111997 values between 0.11 and 2.148 values between 2.14 and 4.164 values between 4.16 and 6.192 values between 8.21 and 10.231 values between 10.23 and 12.261 values between 14.28 and 16.312 values between 16.31 and 18.331 values between 18.33 and 20.351 values between 54.76 and 56.78offnegativepositiveoften

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. iards
  2. ăn
  3. 的家庭
  4. 家喻户晓
  5. 的地图
  6. 戟
  7. Surrey
  8. 媪
  9. .lazy
  10. Anglo

Suppresses

  1. vertex
  2. vertex
  3. vertices
  4. Vertex
  5. Vertex
  6. vertices
  7. Vertices
  8. _vertex
  9. →vertex
  10. (vertex