Neuron Explorer

Qwen3-0.6B

pieces inside split or hyphenated words

A neuron in layer 3 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the middle pieces of words that the tokenizer has broken apart, such as 'zer' in Weezer or 'ed' in 'domed', and on hyphenated links like 'one-and-a-half'. It goes strongly negative on the middle word of 'case-by-case' and 'day-to-day'. Turning it up did not produce a clear change in the writing.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    yura⏎- The Crab Movie, (2010)⏎- Attack of the Crab Monsters (1957)⏎- Crab, by Weezer⏎- The Crab Song, by Faith No More⏎- Crazy Joe, from

  2. #2 peak

    will ask you to walk on our treadmill for up to 15 minutes. You are free to rest at any time. The whole process may take between one and two hours.⏎What is the most important piece of technology used in the

  3. #3 peak

    these in the ground and they will rot away)⏎- plastic containers – these also are takeout containers from a deli and they have the bonus of a domed clear plastic lid

  4. #4 peak

    , and joining the three zones of the canvas: the still life of melons and leaves on a gridded pavement, bottom left; the architecture with domed marabout, top left; and the figures, at right. Next to a

  5. #5 peak

    Auschwitz-Birkenau extermination camp, the last extermination center still functioning under the Nazis. They found 7,650 survivors. Some one-and-a-half million people were murdered there, including 960,00

  6. #6 peak

    .), W: 8 7/8 in. (22.5 cm.)|⏎|Location||Alte Pinakothek, Bayerische Staatsgemä

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 a 2000.100000000000000000000000000000000000000

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

    .⏎Different elements have different melting points, so furnace temperatures are adjusted on a case-by-case basis.⏎In metal parts manufacturing, melting is used for casting -- where the

  2. #2 peak

    to predict what the teacher has written! When they have finished writing their sentence, they stand up. Once all are standing, you begin to read your sentence word by word. Ask them to sit down as soon as one of their words is not

  3. #3 peak

    , let’s start off with a simple question: What is TVOC?⏎Air pollution is an “invisible threat” that most people don’t consider in their day-to-day lives, yet it is an issue that affects people around the world in substantial

  4. #4 peak

    their decision making and strategic planning.⏎• Systematically assess the financial viability of their functions, plans and projects.⏎• Manage and control their budget effectively on a day-to-day basis.⏎Introduction, Course outline, Expectations and Takeaways⏎• Understanding

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. L2:345 weakly excites
  2. L2:2008 weakly suppresses
  3. L2:2692 weakly excites
  4. L1:2688 weakly suppresses
  5. L2:84 weakly excites
  6. L2:111 weakly suppresses
  7. L2:665 weakly excites
  8. L0:178 weakly suppresses
  9. L1:3064 weakly excites
  10. L2:1657 weakly excites
  11. L2:1334 weakly suppresses
  12. L0:1321 weakly suppresses
  13. L0:1224 weakly suppresses
  14. L2:92 weakly excites
  15. L1:1464 weakly excites
  16. L2:302 weakly excites
  17. L1:118 weakly excites
  18. L2:36 weakly excites
  19. L0:540 weakly suppresses
  20. L1:257 weakly suppresses
  21. L1:2922 weakly excites
  22. L1:989 weakly excites
  23. L1:785 weakly suppresses
  24. L1:320 weakly excites

Feeds later layers

  1. L7:1014 suppresses
  2. L8:448 suppresses
  3. L4:723 suppresses
  4. L9:1449 suppresses
  5. L18:2136 suppresses
  6. L4:760 suppresses
  7. L4:1941 suppresses
  8. L5:640 suppresses
  9. L7:42 weakly suppresses
  10. L8:11 weakly suppresses
  11. L7:38 weakly suppresses
  12. L7:836 weakly suppresses
  13. L4:2392 weakly suppresses
  14. L14:917 weakly suppresses
  15. L4:2209 weakly suppresses
  16. L6:1550 weakly suppresses
  17. L6:923 weakly suppresses
  18. L12:51 weakly excites
  19. L11:2292 weakly suppresses 'than' in comparisons
  20. L5:245 weakly suppresses
  21. L17:655 weakly suppresses
  22. L7:1223 weakly suppresses "and" joining two decades or times
  23. L18:2354 weakly suppresses
  24. L17:465 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 -4.04 and -3.901 values between -3.48 and -3.341 values between -2.50 and -2.361 values between -1.10 and -0.961 values between -0.96 and -0.821 values between -0.68 and -0.5410 values between -0.54 and -0.4076 values between -0.40 and -0.26571 values between -0.26 and -0.122046 values between -0.12 and 0.02991 values between 0.02 and 0.16134 values between 0.16 and 0.3044 values between 0.30 and 0.4416 values between 0.44 and 0.588 values between 0.58 and 0.723 values between 0.72 and 0.861 values between 1.00 and 1.141 values between 1.14 and 1.281 values between 1.42 and 1.56offnegativepositiveoften

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. iners
  2. issant
  3. ifting
  4. uisse
  5. riel
  6. liest
  7. iene
  8. izzer
  9. oured
  10. itious

Suppresses

  1. 乍
  2. ...)⏎⏎
  3. 田
  4. GTK
  5. ock
  6. OB
  7. 专栏
  8. block
  9. ...)⏎
  10. pop