Neuron Explorer

Qwen3-0.6B

comma before an 'especially' aside

A neuron in layer 26 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 a comma that comes right before a side remark like 'especially' or 'particularly'. When it was pushed down the model started repeating the word 'especially' over and over, which suggests it normally holds that word back.

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

    a few of the many ways to use edtech:⏎1. Conquering Teacher Shortage and Limited Course Offerings.⏎Finding qualified teachers can be a challenge, especially at this point in the year, but that doesn’t have to stop you

  2. #2 peak

    krit drama). Nandi is the name of Lord Siva’s brisha vahana (bull vehicle). Nandi bull is widely worshipped in the South, particularly in Karnataka. As a matter of fact, the sculpture of Nandi bull

  3. #3 peak

    ism. Some of the most famous ukiyo-e artists include Katsushika Hokusai,<|endoftext|>Weight loss can be a major struggle for some people, so much so that they turn to gastric bypass surgery to help them achieve their weight

  4. #4 peak

    <|endoftext|>Tips for helping children behave⏎(especially in public)⏎I thought it was time to write down some of the tips I’ve acquired that concern children and behavior, as I’ve learned what I have about the brain and how it works. What

  5. #5 peak

    have previously been reported in Europe and India and most recently in Chile. In Washington State and Oregon, Phacidiopycnis rot is common on pears, whereas speck rot is more common on apples in Washington State. Recent surveys have

  6. #6 peak

    - Check the privacy settings and think about who you really want to have access to your personal information.⏎- Be careful about what personal information you put on the internet, these details can be used by scammers to guess your passwords or to commit fraud

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: especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially especially

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

    of a major collapse in agricultural production in poor countries. This will increase malnutrition already at elevated levels in these countries.⏎Sr. Bénédicte Kikwengi, SNDdeN works with her students at the high school in

  2. #2 peak

    , including the formation of placenta which develops from the same early embryo.⏎“The finding was serendipitous and at the same time a very important addition to our understanding of

  3. #3 peak

    agonists (Allbuterol/Salbutamol, Proventil, Ventolin) and theophylline (Theodur, QTSR) work by opening up the bronchial tubes. They may relieve

  4. #4 peak

    in the country just over a decade ago that has revolutionised African banking, and given access to financial services to millions of the poorest people for the first time. M-Pesa<|endoftext|>It has been said that that major historic events are often triggered by

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. L19:1598 strongly suppresses comma before a qualifying clause
  2. L20:1519 strongly excites
  3. L17:118 excites comma after a clause ending
  4. L19:1647 suppresses
  5. L16:240 weakly excites comma after 'sometimes' or 'the truth'
  6. L23:1648 weakly excites
  7. L18:623 weakly suppresses
  8. L14:35 weakly suppresses
  9. L20:2184 weakly excites
  10. L4:1828 weakly excites
  11. L12:878 weakly suppresses
  12. L23:720 weakly suppresses
  13. L16:66 weakly suppresses
  14. L25:116 weakly suppresses
  15. L5:91 weakly excites
  16. L24:2554 weakly excites
  17. L24:8 weakly suppresses
  18. L16:2566 weakly excites
  19. L16:1273 weakly suppresses
  20. L22:592 weakly excites
  21. L10:107 weakly excites unclear: spots where text was removed
  22. L9:2376 weakly excites
  23. L20:476 weakly excites commas that lead into a continuation
  24. L12:1657 weakly excites

Feeds later layers

  1. L27:2047 strongly suppresses
  2. L27:1498 weakly suppresses
  3. L27:1113 weakly excites
  4. L27:997 weakly suppresses
  5. L27:1863 weakly suppresses
  6. L27:1748 weakly suppresses
  7. L27:930 weakly excites
  8. L27:1554 weakly excites
  9. L27:2086 weakly excites
  10. L27:1083 weakly excites
  11. L27:299 weakly excites
  12. L27:990 weakly excites
  13. L27:500 weakly suppresses
  14. L27:1630 weakly excites
  15. L27:2431 weakly excites
  16. L27:1835 weakly excites
  17. L27:244 weakly suppresses
  18. L27:1162 weakly suppresses
  19. L27:1946 weakly excites the letter z inside words
  20. L27:226 weakly suppresses
  21. L27:2649 weakly excites
  22. L27:145 weakly excites dots inside web addresses
  23. L27:18 weakly excites
  24. L27:2611 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 -14.55 and -13.902 values between -7.47 and -6.833 values between -6.19 and -5.544 values between -5.54 and -4.902 values between -4.90 and -4.265 values between -4.26 and -3.614 values between -3.61 and -2.975 values between -2.97 and -2.3315 values between -2.33 and -1.6830 values between -1.68 and -1.0490 values between -1.04 and -0.402434 values between -0.40 and 0.251199 values between 0.25 and 0.8974 values between 0.89 and 1.5311 values between 1.53 and 2.184 values between 2.18 and 2.825 values between 2.82 and 3.464 values between 3.46 and 4.111 values between 4.11 and 4.752 values between 4.75 and 5.393 values between 5.39 and 6.031 values between 6.03 and 6.681 values between 6.68 and 7.323 values between 7.32 and 7.962 values between 7.96 and 8.612 values between 10.54 and 11.18offnegativepositiveoften

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. _hz
  2. lộ
  3. rawl
  4. asmus
  5. 售后
  6. 两地
  7. modity
  8. 瀛
  9. пад
  10. 二代

Suppresses

  1. especially
  2. particularly
  3. particularly
  4. especially
  5. 特别是
  6. 尤其是
  7. 尤其
  8. 特别是在
  9. Especially
  10. 尤其是在