Neuron Explorer

Qwen3-0.6B

unclear: mid-word pieces of long 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

The neuron fires on pieces in the middle of longer words that are split into fragments, like 'iff' in 'Cliffhangers' or 'iz' in 'Janissaries'. It is most strongly turned off on the 'f' in 'liquefied', a word it may partly recognise. Steering outputs did not show a clear theme.

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

    playing with blocks, the breaking down is as important as the building up. That's the type of experience you want to give them in games you design.⏎Cliffhangers and humor⏎The next age group is in to chapter books and tends

  2. #2 peak

    plasma for film, foil, paper, foam, nonwoven and woven substrate applications has had a significant impact on the surfaces of these materials at considerably lower capitalization costs.⏎High-density atmospheric plasma represents a new generation of surface-treatment technology.

  3. #3 peak

    last letters, in any order. If the selected letters match an item in the list, the letters are highlighted and the item is crossed out. Spaces, capitalization and diàçritîcš are ignored. Once the list is cleared

  4. #4 peak

    ing a tribute in children from Christian subjects. The boys were trained to become soldiers and administrators. As soldiers they filled the ranks of the infantry, called the Janizaries (also spelled Janissaries), the most fearsome military force in Europe

  5. #5 peak

    usion-against-the-agri-intellectuals/⏎Leave a ReplyWant to join the discussion?⏎Feel free to contribute!<|endoftext|>Tree Roots Can Wreak Havoc on Pipes — How to Best Prevent Pipe Damage⏎Tree roots craving water will naturally

  6. #6 peak

    ," said Nancy Riestenberg, School Climate Specialist at the Department of Education.⏎Riestenberg joined members of the task force to speak with students from Bemidji High School and students from Schoolcraft to discuss their concerns about bullying and ways

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 way I have been thinking about the problem of the way I have been thinking about the problem of the way I have been thinking about the problem of the way I have been thinking about the problem of the way I have been thinking about the

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

    009, Waste Management, Inc. officially opened the high-tech high Btu fuel plant. The LFG-to-LNG facility converts landfill gas into liquefied natural gas that will fuel 300 garbage trucks operated by Waste Management

  2. #2 peak

    along with some additional information.⏎How is an Acetylene Cylinder Different from Other Compressed Gases?⏎Compressed gases are mainly distinguished into three types – liquefied, gaseous, and dissolved. Acetylene is one of the

  3. #3 peak

    energy needs. For the majority of off-grid operations, this means a choice between oil, LPG (liquefied petroleum gas) or LNG (liquefied natural gas) for high-volume commercial

  4. #4 peak

    million cubic feet of landfill gas, yielding 13,000 gallons of liquefied natural gas to run the garbage trucks and replacing the need for about 2

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. L1:2361 suppresses
  2. L1:951 excites
  3. L2:1710 weakly suppresses
  4. L2:2900 weakly suppresses
  5. L1:2538 weakly excites
  6. L0:1224 weakly suppresses
  7. L2:549 weakly suppresses
  8. L1:2800 weakly excites
  9. L2:49 weakly suppresses
  10. L0:178 weakly suppresses
  11. L0:1819 weakly suppresses
  12. L1:2191 weakly excites
  13. L2:1036 weakly excites
  14. L2:2356 weakly excites
  15. L1:87 weakly suppresses
  16. L2:493 weakly suppresses
  17. L2:2140 weakly suppresses
  18. L1:3003 weakly suppresses
  19. L2:1321 weakly suppresses
  20. L1:622 weakly excites
  21. L0:1301 weakly suppresses
  22. L1:2707 weakly excites
  23. L1:1378 weakly excites
  24. L1:1857 weakly suppresses

Feeds later layers

  1. L4:2505 suppresses
  2. L6:2140 excites
  3. L6:322 suppresses
  4. L18:1893 suppresses
  5. L4:2188 suppresses
  6. L4:528 suppresses
  7. L4:253 weakly suppresses
  8. L9:413 weakly excites
  9. L6:31 weakly suppresses
  10. L17:1651 weakly excites
  11. L7:207 weakly suppresses
  12. L6:1721 weakly suppresses
  13. L12:2394 weakly suppresses
  14. L15:2431 weakly suppresses
  15. L6:32 weakly excites
  16. L26:795 weakly excites
  17. L19:596 weakly suppresses
  18. L7:2642 weakly excites
  19. L11:624 weakly excites
  20. L11:870 weakly suppresses
  21. L4:178 weakly excites
  22. L13:1078 weakly suppresses
  23. L23:467 weakly suppresses
  24. L7:132 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 -2.07 and -2.001 values between -2.00 and -1.921 values between -1.55 and -1.481 values between -0.82 and -0.741 values between -0.67 and -0.602 values between -0.60 and -0.521 values between -0.52 and -0.454 values between -0.45 and -0.386 values between -0.38 and -0.3039 values between -0.30 and -0.2390 values between -0.23 and -0.16328 values between -0.16 and -0.081184 values between -0.08 and -0.011446 values between -0.01 and 0.07515 values between 0.07 and 0.14188 values between 0.14 and 0.2168 values between 0.21 and 0.2920 values between 0.29 and 0.367 values between 0.36 and 0.432 values between 0.43 and 0.512 values between 0.58 and 0.661 values between 0.80 and 0.88offnegativepositiveoften

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. estro
  2. enko
  3. stantial
  4. seau
  5. odable
  6. abar
  7. oproject
  8. odate
  9. oteca
  10. otch

Suppresses

  1. 躺着
  2. ignite
  3. leaf
  4. 化
  5. rug
  6. 烟
  7. публи
  8. REQUEST
  9. ving
  10. tabpanel