Neuron Explorer

Qwen3-0.6B

end-of-text marker between documents

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 hidden end-of-document marker, the point where one web page or text stops and an unrelated one starts. It is nearly silent on ordinary words. Turning it up did not produce any clear change in the written 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

    environmental variables, or logging properties⏎- Monitor Spark using the web UIs, metrics, and external instrumentation⏎- Understand performance tuning considerations⏎- This course is<|endoftext|>Class 1M have been learning the story of Snow White and the Seven D

  2. #2 peak

    for students to individualize the map with images, colors, background, and font. As Miles ‘23 noted, this<|endoftext|>The Bill states that the Constitution of India is dynamic in nature, with the ability

  3. #3 peak

    a writing folder as a<|endoftext|>This page provides information about the questions in Sundanese, also called the interrog

  4. #4 peak

    The most popular toolbars that are used for AutoCAD 2D drawings are the DRAW, MODIFY, and OBJECT PROPERTIES toolbars. Select these toolbars<|endoftext|>February 14, 2007 - Scientists Elucidate the

  5. #5 peak

    reorganize them so that the memory access runs, the R2 operation is held in an Instruction Pool pending completion of the first operation, and R5 and<|endoftext|>ATLANTA (AP) — A recent study shows one in four high school

  6. #6 peak

    D model. Custom-built software then<|endoftext|>August 20, 2012⏎- coddle (

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: how to both, as the students, to the rest, to the students, as the adults, the students, as the institution, as a how to, as, as... a / / / / / S / / / / T

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

    up the data directory that the installed server uses, but also create a pristine “template” data directory named⏎dataunder the installation directory. After an installation has been performed using an MSI package, the template data directory can be copied to set up

  2. #2 peak

    WAS burning !! The relays look like this. Gas-fired models, on the other hand, have a burner at the base of the water heater. I have been using the oven for quite some time with just the broiler

  3. #3 peak

    Core State Standards for states in which they have been adopted. If a state does not appear in the drop-down, CCSS alignments are forthcoming.⏎This lesson has been aligned to standards in the following states

  4. #4 peak

    continue maintenance shots for three to five years.⏎How well do these shots actually work?⏎As with any treatment, effectiveness will vary from person to person. They have been shown to aid in the decrease of symptoms for many different types of allergies.⏎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. L6:449 strongly suppresses
  2. L16:37 strongly suppresses
  3. L5:601 strongly excites
  4. L17:1554 suppresses
  5. L16:478 suppresses
  6. L14:348 suppresses
  7. L16:376 weakly suppresses 'or' joining alternatives
  8. L16:454 weakly excites
  9. L17:300 weakly excites
  10. L5:1896 weakly excites
  11. L3:1480 weakly suppresses
  12. L8:14 weakly excites
  13. L2:469 weakly suppresses
  14. L15:727 weakly excites
  15. L16:112 weakly excites
  16. L15:81 weakly suppresses line breaks inside running text
  17. L6:401 weakly suppresses
  18. L10:32 weakly excites
  19. L16:170 weakly suppresses
  20. L6:359 weakly excites
  21. L0:1746 weakly suppresses
  22. L13:747 weakly excites
  23. L2:1710 weakly suppresses
  24. L3:2484 weakly excites

Feeds later layers

  1. L21:1323 strongly suppresses just before a document ends
  2. L25:436 excites
  3. L27:1300 excites
  4. L22:230 excites
  5. L19:732 excites unclear: end of document token
  6. L24:130 excites line break at the end of a line
  7. L26:603 weakly excites
  8. L20:96 weakly suppresses
  9. L20:1260 weakly suppresses
  10. L19:2591 weakly suppresses
  11. L20:2858 weakly excites
  12. L22:2895 weakly suppresses
  13. L24:1577 weakly suppresses
  14. L20:1391 weakly excites
  15. L21:1767 weakly suppresses
  16. L27:2708 weakly excites
  17. L23:1946 weakly suppresses
  18. L26:145 weakly excites
  19. L22:2951 weakly excites
  20. L19:2811 weakly suppresses
  21. L24:1608 weakly suppresses
  22. L22:3013 weakly excites
  23. L22:133 weakly suppresses
  24. L24:2174 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.

1 values between -12.77 and -12.272 values between -5.28 and -4.784 values between -2.28 and -1.7843 values between -1.78 and -1.2892 values between -1.28 and -0.78492 values between -0.78 and -0.293121 values between -0.29 and 0.21139 values between 0.21 and 0.714 values between 0.71 and 1.215 values between 1.21 and 1.711 values between 1.71 and 2.211 values between 2.71 and 3.211 values between 3.21 and 3.711 values between 5.21 and 5.711 values between 6.70 and 7.20offnegativepositiveoften

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. arga
  2. esto
  3. _world
  4. 帐
  5. казать
  6. 这才
  7. -contrib
  8. front
  9. deviation
  10. ESİ

Suppresses

  1. artner
  2. ạch
  3. burned
  4. dux
  5. éta
  6. .colorbar
  7. uitar
  8. ammen
  9. berger
  10. acades