Neuron Explorer

Qwen3-0.6B

the word 'even' after a clause

A neuron in layer 5 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 the word 'even' when it follows a phrase and adds a surprising extra, as in 'removing obstacles even if we don't know it'. Turning it up gave unrelated text.

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

    , but a means to an end, the end being spiritual solitude. In spiritual solitude the soul is alone with God, attentive to him in preference to creatures, even though it has to deal with

  2. #2 peak

    , whether it turns out that the butler did it, or if it was Colonel Mustard in the library with a candlestick.⏎But I love scientific mysteries even more.⏎Recently, scientists doing research at Fermi National Accelerator Laboratory, or

  3. #3 peak

    whatever comes as if it’s our greatest happiness. Say, “It’s just like that,” and keep going.⏎The Friend is generous and wise, removing obstacles even if we don’t know it, providing our daily portion from His granaries in

  4. #4 peak

    bad qualities within all living beings that hinder our path to Buddhahood, such as greed, hatred, and pride. Buddhists have compassion for all beings, even the demons, who must be subdued and brought back into the fold. Every being

  5. #5 peak

    that sort of came along and stole the thunder of the payphone?” Rocca asked.⏎“It stole its relevance. And it made communication so simple and so trivial even

  6. #6 peak

    Shem Rah⏎- Listening to the above⏎- Believing the above⏎- Any form of communication about a fellow Jew that is⏎derogatory, even if it does not cause him/her any harm.⏎- Any form of communication that

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 process of the development of the story, and the story of the characters in the story of the game. The development of the game is in the form of a single entity. I will now present the story of the game in the form

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

    the world. The factbook gives a two- to three-page summary of the demographics, location, telecommunications capacity, government, industry, military capability, etc, of all US-recognized countries and territories in the world.⏎As The World Factbook

  2. #2 peak

    <|endoftext|>Planning is the key to successful grove establishment, maintenance, and production. Developing a detailed infrastructure description and plan, cultural program, and financial and marketing plan for a new or existing grove with a new fruit crop will save you time and

  3. #3 peak

    articles and related images, headlines, ads, and title of the publication.⏎- Brainstorm about what you would like your newspaper to feature—

  4. #4 peak

    . Our General Education program is designed to provide both depth and breadth of knowledge, and coupled with courses in your major and the co-curricular activities you choose, to help you develop the attributes of an educated person. The UMass Dartmouth Commit

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. L3:938 excites
  2. L4:153 suppresses
  3. L4:436 weakly suppresses
  4. L4:91 weakly excites
  5. L1:552 weakly excites
  6. L2:339 weakly excites
  7. L2:695 weakly suppresses
  8. L0:254 weakly suppresses
  9. L3:94 weakly suppresses
  10. L0:341 weakly suppresses
  11. L2:1402 weakly suppresses
  12. L2:218 weakly excites
  13. L4:77 weakly suppresses
  14. L1:1481 weakly suppresses
  15. L2:62 weakly excites
  16. L2:1528 weakly suppresses
  17. L2:2303 weakly excites
  18. L3:73 weakly suppresses
  19. L4:276 weakly excites
  20. L3:2214 weakly suppresses period after a capital initial
  21. L4:248 weakly excites
  22. L1:1197 weakly suppresses
  23. L0:2196 weakly excites
  24. L0:2110 weakly excites

Feeds later layers

  1. L13:2729 excites
  2. L26:2238 weakly excites
  3. L6:104 weakly excites
  4. L7:821 weakly excites
  5. L12:1388 weakly excites day number after a month name
  6. L6:1287 weakly suppresses
  7. L6:70 weakly suppresses
  8. L7:578 weakly excites 'and' in lists of three
  9. L10:703 weakly suppresses
  10. L11:477 weakly suppresses
  11. L7:761 weakly excites
  12. L10:1103 weakly excites
  13. L12:500 weakly excites
  14. L10:1284 weakly excites
  15. L7:270 weakly excites
  16. L8:2812 weakly excites
  17. L13:295 weakly suppresses
  18. L10:2645 weakly excites
  19. L8:101 weakly suppresses unclear: commas and 'or' in lists
  20. L8:349 weakly excites 'of' inside title-like noun phrases
  21. L7:329 weakly suppresses
  22. L22:1662 weakly excites
  23. L11:426 weakly suppresses
  24. L23:1395 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.

1 values between -2.05 and -1.941 values between -1.41 and -1.302 values between -1.30 and -1.191 values between -1.19 and -1.092 values between -0.98 and -0.871 values between -0.87 and -0.775 values between -0.77 and -0.669 values between -0.66 and -0.5516 values between -0.55 and -0.4533 values between -0.45 and -0.3476 values between -0.34 and -0.23249 values between -0.23 and -0.13958 values between -0.13 and -0.021788 values between -0.02 and 0.09447 values between 0.09 and 0.20168 values between 0.20 and 0.3082 values between 0.30 and 0.4130 values between 0.41 and 0.5213 values between 0.52 and 0.629 values between 0.62 and 0.738 values between 0.73 and 0.844 values between 0.84 and 0.941 values between 0.94 and 1.051 values between 1.26 and 1.371 values between 1.48 and 1.581 values between 1.69 and 1.801 values between 2.12 and 2.23offnegativepositiveoften

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. anny
  2. edom
  3. son
  4. venes
  5. millennium
  6. INDER
  7. erno
  8. shepherd
  9. -Za
  10. anniversary

Suppresses

  1. ... ⏎
  2. ..."⏎
  3. ...'⏎
  4. 归属于
  5. ęk
  6. ....⏎
  7. ####⏎
  8. 创建
  9. ...⏎
  10. ...";⏎