Neuron Explorer

Qwen3-0.6B

the word 'would' before a verb

A neuron in layer 17 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 lights up on the word "would" when it comes before the verb that follows, as in "it would be better" or "what would you do." When it was turned up, the model's text wandered to unrelated topics, so the effect on the writing is not clear.

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

    / Guide because after using our Pdf solutions you will never find need of them. You will get all the Pdf files at the end of this post, so we would like to request you to read full article.⏎|Maharashtra state Board||

  2. #2 peak

    done this way because it’s easy for normal Drupal users to control on a shared host. That’s not to say it isn’t a valid solution just that it would typically be faster to cache to a file.⏎Anyway, perhaps one of the most

  3. #3 peak

    using illeism: “One thing I didn’t want to do was make an emotional decision … I wanted to do what was best for LeBron James … what would make him happy.”⏎James was lampooned for speaking this way and accused of

  4. #4 peak

    derivation of Kirchhoff's voltage law.⏎Batteries do make a good example for this simply because they are current sources with completely isolated grounds. This example would be

  5. #5 peak

    Plan⏎Make A Plan⏎Disaster can strike quickly and without warning. It can force you to evacuate your neighborhood or confine you to your home. What would you do if basic services were cut off? Local officials and relief workers will be

  6. #6 peak

    still in the top ten. Most of such calls are followed by specific attribute access.⏎Why use anecdotal evidence? I don't doubt the numbers, but it would be⏎better to use the standard library

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: I have to to do something to the end of a to a to a to... I am to a to a to...... (I am to a to a to...) I have to to the end of the end of the end

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

    , he began travelling to the USA and Europe where he received more commendations for his artistic gift.⏎His grandest work is the Symphony of Light, which was supposed to

  2. #2 peak

    . It is deeply influenced by the country's culture, religion, and historical events, and has evolved over time to reflect changing artistic styles and techniques. In this blog post, we'll explore the fascinating world of

  3. #3 peak

    obvious; we hope some of the technical information is useful. For hearing interviewers, consider who is on your interview team. Collaboration with Deaf people–cinematographers, editors, Certified Deaf Interpreters (CDIs), production assistants

  4. #4 peak

    use them with your low beams during inclement weather.⏎<|endoftext|>The equipment of the standing army is thoroughly modern, the infantry being provided with Mauser rifles and the artillery with Kr

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. L2:1436 strongly excites
  2. L16:120 strongly excites
  3. L6:299 strongly excites
  4. L0:1872 suppresses
  5. L16:100 excites
  6. L15:89 excites
  7. L15:114 excites
  8. L15:792 suppresses
  9. L16:296 suppresses
  10. L15:78 excites
  11. L15:115 excites
  12. L6:840 weakly suppresses
  13. L15:1587 weakly excites
  14. L3:1697 weakly excites
  15. L15:82 weakly suppresses
  16. L15:1404 weakly suppresses
  17. L15:96 weakly suppresses
  18. L16:28 weakly suppresses
  19. L15:129 weakly suppresses
  20. L15:201 weakly excites
  21. L0:983 weakly suppresses "similar" after "are"
  22. L15:451 weakly excites
  23. L15:1616 weakly excites word fragments mid-word
  24. L16:79 weakly suppresses

Feeds later layers

  1. L18:719 strongly excites
  2. L26:466 excites
  3. L21:108 suppresses
  4. L22:1848 excites
  5. L18:924 suppresses
  6. L22:2361 excites
  7. L27:2265 excites
  8. L18:1442 suppresses unclear: word fragments and digits inside numbers
  9. L26:1989 suppresses
  10. L22:2970 suppresses
  11. L19:1258 suppresses
  12. L22:2274 suppresses
  13. L21:2617 suppresses
  14. L20:2885 suppresses
  15. L26:1760 suppresses
  16. L26:2006 excites
  17. L20:2681 excites
  18. L18:3058 excites
  19. L20:626 suppresses
  20. L18:872 excites the digit 8 in lists and numbers
  21. L26:1595 suppresses
  22. L21:2715 weakly excites
  23. L26:2364 weakly excites
  24. L19:3002 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 -3.57 and -3.041 values between -2.52 and -1.995 values between -1.99 and -1.47913 values between -1.47 and -0.951895 values between -0.95 and -0.42600 values between -0.42 and 0.11111 values between 0.11 and 0.6398 values between 0.63 and 1.1558 values between 1.15 and 1.6848 values between 1.68 and 2.2040 values between 2.20 and 2.7334 values between 2.73 and 3.2527 values between 3.25 and 3.7818 values between 3.78 and 4.309 values between 4.30 and 4.8310 values between 4.83 and 5.359 values between 5.35 and 5.886 values between 5.88 and 6.409 values between 6.40 and 6.931 values between 6.93 and 7.454 values between 7.45 and 7.981 values between 7.98 and 8.503 values between 9.03 and 9.551 values between 9.55 and 10.081 values between 10.60 and 11.131 values between 11.13 and 11.651 values between 12.17 and 12.702 values between 16.90 and 17.42offnegativepositiveoften

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. 불구하고
  2. ट
  3. forces
  4. gne
  5. 峭
  6. ailable
  7. logen
  8. harbour
  9. →ev
  10. driv

Suppresses

  1. ".$
  2. arity
  3. credited
  4. ></
  5. overtime
  6. esimal
  7. ".$_
  8. ><
  9. fact
  10. �认