Neuron Explorer

Qwen3-0.6B

unclear: document boundaries and digits

A neuron in layer 10 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 is hard to pin down. It fires most on the hidden end-of-document marker that web pages sometimes contain, and on some digits inside ISBN numbers and web addresses. Turning it up gave no 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

    :<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Aim High' printed from

  2. #2 peak

    :<|endoftext|>The hamstring muscles – the biceps femoris, semitendinosus,

  3. #3 peak

    be the owner of a collection of Late Byzantine Mosaic Icons… but there is no proof for this. https://www.academia.edu/39903340/Displaying_an_Icon_The_Mosaic_

  4. #4 peak

    0240.<|endoftext|>By 2003, plenty of data suggested that pesticide drift was harming

  5. #5 peak

    article helpful?<|endoftext|>Given that this is the first U.S. presidential election since apps have made their

  6. #6 peak

    Ehrlichia canis in naturally infected ticks and sequence analysis of p30-10 of E canis from diverse geographic regions.(opens in new tab)⏎Journal of Clinical Microbiology⏎Suleyman FelekY Rik

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 do I get to the point where I can express my thought about the topic of science. What is the level of expression? I need to be able to express my thought via any medium and form of expression. Level of expression. I

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

    roll, Lewis. Alice in Wonderland. Macmillan Children's Books, 2015. 208 pp., ISBN 9781447279990<|endoftext|>It won’t be wrong to

  2. #2 peak

    American Studies scholar Jean O’Brien: Drop Dead<|endoftext|>Note: Supplemental materials are not guaranteed with Rental or Used book purchases.⏎- ISBN: 9781908049490 | 19080

  3. #3 peak

    HyperParameters:⏎Gradient descent takes small steps to reduce the loss of a model. Gradient descent

  4. #4 peak

    much more!⏎Discover the world of gorillas and other primates their remarkable intelligence and social behavior⏎66 pages; ISBN 9780756668082⏎, or download in⏎More

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. L9:1115 weakly excites
  2. L9:1532 weakly excites
  3. L4:1183 weakly suppresses
  4. L8:1874 weakly suppresses
  5. L6:1057 weakly excites
  6. L9:214 weakly excites
  7. L7:826 weakly suppresses
  8. L8:346 weakly suppresses
  9. L0:2648 weakly excites
  10. L7:1008 weakly excites
  11. L3:60 weakly suppresses
  12. L3:2722 weakly suppresses
  13. L9:74 weakly excites
  14. L4:2166 weakly excites
  15. L8:327 weakly suppresses
  16. L3:702 weakly excites
  17. L8:2926 weakly suppresses
  18. L9:1489 weakly excites
  19. L7:170 weakly suppresses
  20. L3:109 weakly suppresses
  21. L6:2157 weakly excites
  22. L8:2200 weakly excites
  23. L8:447 weakly suppresses
  24. L0:467 weakly excites

Feeds later layers

  1. L11:1047 suppresses pieces of software and library names
  2. L13:2523 suppresses
  3. L14:1248 suppresses
  4. L12:71 weakly suppresses
  5. L11:1072 weakly suppresses
  6. L19:913 weakly suppresses
  7. L14:2069 weakly excites
  8. L19:1717 weakly suppresses
  9. L25:412 weakly suppresses
  10. L16:910 weakly suppresses
  11. L20:1284 weakly suppresses
  12. L16:2782 weakly excites
  13. L16:2579 weakly suppresses
  14. L11:2326 weakly suppresses
  15. L11:2295 weakly suppresses
  16. L11:502 weakly suppresses
  17. L27:952 weakly excites
  18. L11:2535 weakly suppresses
  19. L16:73 weakly suppresses
  20. L13:1354 weakly suppresses
  21. L20:1904 weakly excites
  22. L16:46 weakly suppresses
  23. L16:86 weakly suppresses
  24. L21:1550 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.

1 values between -0.66 and -0.604 values between -0.60 and -0.543 values between -0.54 and -0.484 values between -0.48 and -0.438 values between -0.43 and -0.3714 values between -0.37 and -0.3132 values between -0.31 and -0.2564 values between -0.25 and -0.19162 values between -0.19 and -0.13344 values between -0.13 and -0.08865 values between -0.08 and -0.021150 values between -0.02 and 0.04659 values between 0.04 and 0.10265 values between 0.10 and 0.16166 values between 0.16 and 0.2284 values between 0.22 and 0.2750 values between 0.27 and 0.3318 values between 0.33 and 0.397 values between 0.39 and 0.456 values between 0.45 and 0.511 values between 0.62 and 0.681 values between 1.61 and 1.67offnegativepositiveoften

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. ophon
  2. ธร
  3. ACP
  4. >Main
  5. vanished
  6. onPage
  7. çi
  8. aji
  9. Magnet
  10. 込まれ

Suppresses

  1. {}.
  2. nte
  3. 纹
  4. ngle
  5. otypes
  6. 骤
  7. regation
  8. 千万
  9. quires
  10. 类型