Neuron Explorer

Qwen3-0.6B

blank spot before a number

A neuron in layer 18 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 empty gap right before a number, like the space in 'bundles of ___3D' or 'in ___3 different ways'. It seems to be waiting for a digit. Turning it up or down just made the text wander into broken or repetitive wording.

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

    Using a new printed optics technology developed by Disney Research Pittsburgh and Carnegie Mellon University, Brockmeyer and Poupyrev have designed character eyeballs made from bundles of 3D printed optical fibers.⏎This novel technology, named PAPILLON

  2. #2 peak

    cookies or want to be informed before they are stored on their computer can configure their browser for this purpose.⏎Most of today’s browsers allow the management of cookies in 3 different ways:⏎The cookies are never accepted.⏎For each cookie, the browser

  3. #3 peak

    2021 - Vaccine Types and Their Effectiveness⏎The recent approval of the commercial growing of Golden Rice (GR) based on guidelines set by a 5-interdepartmental circular is a most welcome, most awaited development.⏎Vitamin

  4. #4 peak

    Flight Test Instrumentation Technicians who all have their eyes on the same goal—safety.⏎After all, the last place you want to have any problems is 30,000 feet in the air. So each plane is thoroughly

  5. #5 peak

    is where the exchange media (tiny beads) is contained. Some water softeners come as one unit (mineral tank inside of brine tank, or as 2 separate units.⏎The brine tank should always have salt in order for a

  6. #6 peak

    century when Claude Stairs, a Belgian inventor came up with a great idea. Not many people know that lifts were invented thousands of years ago but stairs are a 13th century invention. Claude Stairs was working on the court of Count

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:Ifapersontwiththefollowingconditiondwiththefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowingconditionhavingthefollowing

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

    "no" "yes"⏎ghci> map yesOrNo [True, False, True]⏎["yes", "no", "yes"]⏎Note that the arguments to⏎bool are in the opposite

  2. #2 peak

    *) memchr (str, 'p', strlen(str));⏎printf ("'p' found at position %d.\n", pch-str+1);⏎printf ("'p' not found.\n");⏎- Compare two blocks of memory (function

  3. #3 peak

    is used to count the numbers of characters in⏎a string, as well as the number of characters in the string variable⏎>>> len("a string") 8 >>> >>> s = "a long string" >>> >>>

  4. #4 peak

    <codeBase> element, which specifies where the runtime can locate an assembly (for more information, see Specifying an Assembly's Location).⏎<codeBase version="188.8.131.52" href="http

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. L14:140 strongly suppresses
  2. L6:344 strongly excites
  3. L16:396 strongly suppresses
  4. L6:42 strongly suppresses
  5. L16:723 strongly suppresses
  6. L6:1550 strongly excites
  7. L9:974 strongly suppresses
  8. L5:2539 excites
  9. L6:917 suppresses comma after a month and day
  10. L4:874 suppresses opening parenthesis before acronyms
  11. L16:22 suppresses
  12. L4:1977 suppresses
  13. L12:2582 suppresses
  14. L4:2129 suppresses
  15. L16:1421 suppresses
  16. L17:15 excites
  17. L14:102 excites
  18. L14:309 excites
  19. L5:2991 suppresses
  20. L13:2204 suppresses
  21. L4:447 suppresses
  22. L4:1129 excites
  23. L14:51 excites colon that introduces a statement
  24. L9:2631 suppresses

Feeds later layers

  1. L24:51 strongly suppresses
  2. L21:29 strongly suppresses
  3. L26:17 strongly suppresses
  4. L19:509 strongly suppresses
  5. L21:230 strongly suppresses
  6. L23:262 strongly suppresses
  7. L24:77 strongly suppresses opening parenthesis after a name
  8. L20:563 strongly suppresses
  9. L25:222 strongly suppresses
  10. L26:240 strongly suppresses
  11. L22:14 suppresses
  12. L25:217 excites
  13. L26:21 excites
  14. L20:59 suppresses
  15. L27:68 suppresses
  16. L27:424 suppresses hyphen inside a compound word
  17. L24:141 suppresses
  18. L25:30 suppresses
  19. L27:11 excites unclear: mixed fragments in lists and pronunciations
  20. L25:304 suppresses
  21. L27:318 suppresses opening parenthesis after a name or term
  22. L27:33 suppresses
  23. L23:80 suppresses blank slot before a year or number
  24. L27:101 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.

1 values between -17.13 and -16.452 values between -11.06 and -10.391 values between -9.71 and -9.045 values between -9.04 and -8.372 values between -8.37 and -7.692 values between -7.69 and -7.023 values between -7.02 and -6.346 values between -6.34 and -5.678 values between -5.67 and -5.0010 values between -5.00 and -4.325 values between -4.32 and -3.6512 values between -3.65 and -2.9812 values between -2.98 and -2.3011 values between -2.30 and -1.6316 values between -1.63 and -0.9520 values between -0.95 and -0.282222 values between -0.28 and 0.401411 values between 0.40 and 1.07148 values between 1.07 and 1.744 values between 1.74 and 2.423 values between 2.42 and 3.091 values between 4.44 and 5.112 values between 5.11 and 5.791 values between 9.15 and 9.83offnegativepositiveoften

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. Le
  2. le
  3. Shadow
  4. re
  5. shadow
  6. accord
  7. EN
  8. Re
  9. dot
  10. Fed

Suppresses

  1. which
  2. Which
  3. 特别是
  4. đ
  5. segue
  6. 尤其是
  7. Indeed
  8. 特殊的
  9. Maximum
  10. constructed