Neuron Explorer

Qwen3-0.6B

word just before an example or elaboration

A neuron in layer 14 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

This neuron fires on the word right before a clarifying detail, like 'most [often]', 'ingredient[:]', or 'liquids, [usually]', where the sentence is about to add an example or explanation. Turning it up made the model drift into odd repeated phrasing, so the link is only loosely supported.

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

    of neutrons—fuel pellets—are placed together. The pellets bombard each other with neutrons, causing fission. Fission happens when a heavier element, most often a form of uranium, is split into smaller pieces, releasing energy. As each

  2. #2 peak

    grading.⏎Next time you hand back that assignment with a grade and comments, sit back and watch the kids. I will almost guarantee they will look at two things - the first will be their own mark and the second will be their neighbour's mark

  3. #3 peak

    a growing list of products marketed to familes that could be harmful. Much like Tide Free and Gentle laundry detergent, it can sometimes contain a cancer-causing ingredient: talcum powder.⏎Babies aren’t the only ones using baby powder. Many

  4. #4 peak

    of a religious institute; or temporarily,k as in a retreat. Physical solitude is not escapism or isolationism, but a means to an end, the end being spiritual solitude. In spiritual solitude the soul is alone with God, attentive to him

  5. #5 peak

    other wet metal objects from rusting into dust prematurely. Why do you need a sacrificial anode? Because, by using an anode made of a metal like zinc, the anode will

  6. #6 peak

    physicians and researchers identified a link in the Brazilian sample population that has been recognized for at least ten years in other countries: the association of two types of psychiatric disorders—depression and anxiety—with a higher risk of

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 "frobenius, the "frobenius, the "fro, the, the, "the, the, the, "fro, "the, "fro, "fro, "fro

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

    easily electricity flows along a certain path through an electrical element. The SI derived unit of conductance is the siemens (formerly referred to as the reciprocal ohm or mho). Oliver Heaviside coined the term in September 188

  2. #2 peak

    subjects, or indeed anyone with an interest in Black and minority ethnic struggles, there are books on all levels of detail, from the broad sweep of historic monographs to the specificity of marginalised group politics, for example Cherokee Women in Crisis, or

  3. #3 peak

    didn’t realize it at the time, but he was already practicing socially responsible investing.⏎What is Socially Responsible Investing?⏎Socially responsible investing (sometimes also called Ethical Investing) is exactly what it sounds like. It’s making sure your

  4. #4 peak

    more likely to be injured.⏎Helmet usage is also on the decline, dropping from 66 percent of motorcyclists wearing helmets in 2011 to only 60 percent in 2012. Head injury is the

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. L6:1567 excites colon introducing a summary point
  2. L13:287 suppresses
  3. L13:948 suppresses
  4. L13:2519 suppresses
  5. L12:1640 excites
  6. L8:749 suppresses
  7. L8:2127 weakly excites
  8. L13:2235 weakly suppresses unclear: metadata labels and closing quotes
  9. L5:563 weakly excites
  10. L6:1348 weakly suppresses
  11. L13:524 weakly suppresses
  12. L10:272 weakly suppresses
  13. L8:1608 weakly suppresses
  14. L11:2662 weakly suppresses
  15. L4:24 weakly excites
  16. L6:1066 weakly suppresses
  17. L3:76 weakly suppresses
  18. L2:2347 weakly suppresses
  19. L13:979 weakly excites
  20. L9:831 weakly excites
  21. L12:1120 weakly suppresses
  22. L13:484 weakly excites
  23. L6:858 weakly excites
  24. L12:434 weakly suppresses

Feeds later layers

  1. L16:1527 suppresses
  2. L15:1299 excites
  3. L17:1002 suppresses
  4. L15:2090 suppresses
  5. L15:1438 weakly suppresses
  6. L17:2482 weakly suppresses
  7. L15:218 weakly suppresses
  8. L16:229 weakly excites
  9. L19:177 weakly suppresses
  10. L18:1143 weakly suppresses
  11. L16:431 weakly excites
  12. L16:2833 weakly suppresses
  13. L17:982 weakly suppresses
  14. L16:611 weakly suppresses
  15. L18:19 weakly excites
  16. L15:370 weakly excites
  17. L15:2787 weakly suppresses
  18. L16:297 weakly excites
  19. L19:2760 weakly suppresses
  20. L16:1548 weakly suppresses
  21. L18:2479 weakly suppresses
  22. L20:1976 weakly suppresses
  23. L16:2296 weakly suppresses
  24. L16:990 weakly 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.

2 values between -1.64 and -1.481 values between -1.48 and -1.334 values between -1.33 and -1.172 values between -1.17 and -1.017 values between -1.01 and -0.866 values between -0.86 and -0.7019 values between -0.70 and -0.5559 values between -0.55 and -0.39233 values between -0.39 and -0.23847 values between -0.23 and -0.082007 values between -0.08 and 0.08479 values between 0.08 and 0.24103 values between 0.24 and 0.3952 values between 0.39 and 0.5528 values between 0.55 and 0.7014 values between 0.70 and 0.868 values between 0.86 and 1.025 values between 1.02 and 1.174 values between 1.17 and 1.333 values between 1.33 and 1.494 values between 1.49 and 1.645 values between 1.64 and 1.804 values between 1.80 and 1.951 values between 1.95 and 2.112 values between 2.11 and 2.272 values between 2.27 and 2.422 values between 2.42 and 2.583 values between 2.58 and 2.741 values between 3.20 and 3.361 values between 4.45 and 4.61offnegativepositiveoften

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. setTimeout
  2. fadeOut
  3. 前锋
  4. assador
  5. entina
  6. 不行
  7. 扫一扫
  8. دور
  9. adders
  10. 深交

Suppresses

  1. cribes
  2. 贡
  3. graded
  4. enity
  5. uced
  6. toned
  7. 、“
  8. itated
  9. ง
  10. ="../../../