Neuron Explorer

Qwen3-0.6B

end of a document marker

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 turns on near the point where one web text ends and an unrelated one begins, often on the marker that separates documents. Turning it up gave text that was off-topic or looped, so it does not show a clear effect on what the model writes.

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

    more resistant than any metal option would be. Using graphite bearings in the mechanical bearing situations makes it possible to explore even more temperature and chemical ranges in these operations.<|endoftext|>Gardenias are sub-tropical evergreen flowering shrubs that bloom in the

  2. #2 peak

    Escherichia coli are investigated by many authors and give the opportunity to perform a thorough comparison of most of the bioinformatics methods used to detect HTs.<|endoftext|>New York City, 06

  3. #3 peak

    in speed regulation range, large in low-speed torque, small in fluctuation, stable in operation, low in noise, high in efficiency and without electromagnetic interference.<|endoftext|>German military cemetery Kharkov, Ukraine⏎During World War II, Khark

  4. #4 peak

    range of sash positions. This range should be determined and marked during fume hood testing. For hoods with face velocity controls, it is imperative to keep<|endoftext|>Interesting Facts about the Blow Ye Winds (extra verses)⏎I want to introduce Blow

  5. #5 peak

    . This would lead to food insecurity, which in turn could lead to social instability, as was already the case in Mozambique in 2010.<|endoftext|>This PBS video focuses on sea level rise in Norfolk, Virginia and how the residents

  6. #6 peak

    the damaged areas.⏎- Maintenance – Cleaning and clearing sewer lines regularly will prevent root growth inside your pipes. As a result, you’ll avoid major sewer repairs.<|endoftext|>Anthropogenic land-use change intensifies the effect of low flows on stream fishes

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: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

    leading to better digestion; however, this argument does not physiologically hold merit since the acid in the stomach inactivates the enzymes once the food is eaten. Plus, the body already makes all the enzymes it needs to digest and absorb food

  2. #2 peak

    and a food supply (such as clothing, wood or drywall).⏎Mould spores and the toxins produced by certain moulds may be harmful to vulnerable people.⏎How could my family and I be exposed?⏎Mould spores

  3. #3 peak

    the first celebration of July 4th⏎Where the majority of recycled glass goes:⏎A. South for the winter⏎B. Back into new bottles and jars⏎C. Roadway sub-base⏎The primary ingredients of a glass container:⏎A

  4. #4 peak

    . .⏎.related to defective perception or wits). TeeMaiR (a dental-shift from⏎DM)means to rise in a cloud see THYME.⏎DIM is traced to Old Norse dimmr (dark) andIEbasedhem (

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. L17:61 excites
  2. L15:2706 excites
  3. L10:65 suppresses
  4. L6:13 suppresses
  5. L15:2567 excites
  6. L17:849 suppresses
  7. L6:7 excites
  8. L17:796 weakly excites
  9. L11:478 weakly excites
  10. L16:709 weakly excites
  11. L2:75 weakly excites
  12. L16:2610 weakly excites
  13. L6:17 weakly excites
  14. L7:303 weakly excites
  15. L8:97 weakly suppresses
  16. L11:17 weakly excites
  17. L17:2272 weakly suppresses
  18. L8:7 weakly suppresses
  19. L14:1266 weakly suppresses
  20. L17:24 weakly suppresses
  21. L7:64 weakly excites
  22. L15:995 weakly excites
  23. L6:8 weakly excites
  24. L5:110 weakly excites

Feeds later layers

  1. L19:2045 suppresses abbreviation pieces in organisation names
  2. L26:26 excites
  3. L19:116 excites
  4. L21:48 excites
  5. L19:2303 suppresses
  6. L19:1038 suppresses
  7. L21:59 excites
  8. L19:423 excites words that set up a following clause
  9. L19:530 suppresses
  10. L19:1172 weakly suppresses
  11. L21:377 weakly suppresses
  12. L20:368 weakly suppresses
  13. L19:1067 weakly suppresses
  14. L22:112 weakly excites
  15. L19:440 weakly suppresses
  16. L19:24 weakly excites
  17. L21:21 weakly excites
  18. L22:8 weakly excites
  19. L19:2362 weakly suppresses
  20. L19:2600 weakly suppresses
  21. L22:238 weakly suppresses
  22. L19:1293 weakly suppresses
  23. L19:26 weakly excites
  24. L22:109 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 -7.51 and -7.023 values between -7.02 and -6.521 values between -6.52 and -6.036 values between -6.03 and -5.538 values between -5.53 and -5.0414 values between -5.04 and -4.5425 values between -4.54 and -4.0538 values between -4.05 and -3.5573 values between -3.55 and -3.06102 values between -3.06 and -2.56145 values between -2.56 and -2.06194 values between -2.06 and -1.57261 values between -1.57 and -1.07319 values between -1.07 and -0.58350 values between -0.58 and -0.08422 values between -0.08 and 0.41388 values between 0.41 and 0.91335 values between 0.91 and 1.40320 values between 1.40 and 1.90249 values between 1.90 and 2.39200 values between 2.39 and 2.89132 values between 2.89 and 3.38100 values between 3.38 and 3.8867 values between 3.88 and 4.3732 values between 4.37 and 4.8741 values between 4.87 and 5.3625 values between 5.36 and 5.8621 values between 5.86 and 6.3516 values between 6.35 and 6.858 values between 6.85 and 7.354 values between 7.35 and 7.842 values between 7.84 and 8.342 values between 8.34 and 8.831 values between 9.33 and 9.821 values between 11.31 and 11.801 values between 11.80 and 12.30offnegativepositiveoften

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. en
  2. le
  3. el
  4. entionPolicy
  5. holds
  6. ob
  7. hold
  8. study
  9. top
  10. y

Suppresses

  1. ɯ
  2. dana
  3. ']}</
  4. —he
  5. IKE
  6. ']*
  7. ards
  8. dda
  9. keeper
  10. IFF