Neuron Explorer

Qwen3-0.6B

mid-word syllables in capitalised names

A neuron in layer 1 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 reacts to the middle pieces of longer names and words, especially ones that start with 'Al', 'Ba' or 'El', such as ALBANIA, Balboa, Albatross and Elbe. When it was turned up, the model's writing did not visibly change in any clear way.

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

    other technologies are used in this course. This course fulfills the graduation requirement for fine and/or practical arts.<|endoftext|>|Albania Table of Contents⏎ALBANIA WAS THE LAST COUNTRY in Eastern Europe during the early 19

  2. #2 peak

    dernière mise à jour : 05/07/2021⏎Ajouter un commentaire<|endoftext|>By DANIEL SUDDEATH⏎NEW ALBANY —⏎Judging by the pleasure clearly shown by the children as they whe

  3. #3 peak

    of The World As It Could Be Tree Dedication Ceremony, students, teachers, and other collaborators on the project reflect on the Universal Declaration of Human Rights. Balboa High School Tree Ceremony from BAYCAT.<|endoftext|>Several images have been returned to

  4. #4 peak

    ceptor Group, his fifth combat victory of the Korean War. Green safely ejected and was rescued from the Yellow Sea by an Air Rescue Service SA-16 Albatross flying boat.⏎Captain Fedorets closed on McConnell from

  5. #5 peak

    The Hobbit”” Smeagol is portrayed as destructive and contemptuous in that he tries to eat Bilbo a fellow hobbit only since he needs something to eat and raw fish are entirely

  6. #6 peak

    , removing the silt from the channel, the dock and the harbor entrance. For example, the entrance of Port of Hamburg located in the estuary of Elbe River in German, the Port of Rotterdam in the Netherlands and the Port of Ant

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 been thinking about lately, and I have been thinking about lately. What do you think about the latest updates on this topic? What have you thought? What have you thought? What do you think about the latest updates on this topic

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

    of Health and Human Services Division of Public Health as well as data from the North Carolina Birth Defects Monitoring Program.<|endoftext|>BARLEY (Hordeum vulgare)⏎Requires six hours or more⏎of strong, direct sun per day.⏎

  2. #2 peak

    artan steel and concrete. None of the statuary of mermen on dolphin and allegorical sculptures are present in today’s Jones Bridge.⏎Puente de Colgante⏎The construction of the Puente Colgante was completed in 1

  3. #3 peak

    Art.⏎There are no reviews yet.<|endoftext|>The first three bridges that spanned across the Pasig River were the Puente Grande, Puente de Colgante and Puente de Ayala.⏎Built in 1632,

  4. #4 peak

    As the name suggest, Puente Colgante was the first suspension bridge to be built in the country. It is known

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. L0:178 suppresses
  2. L0:1224 suppresses
  3. L0:1301 suppresses
  4. L0:512 weakly suppresses
  5. L0:979 weakly excites
  6. L0:718 weakly suppresses
  7. L0:497 weakly suppresses
  8. L0:435 weakly suppresses
  9. L0:1441 weakly excites
  10. L0:1 weakly suppresses
  11. L0:1807 weakly suppresses
  12. L0:2317 weakly excites
  13. L0:1390 weakly excites
  14. L0:210 weakly suppresses
  15. L0:2682 weakly excites
  16. L0:1634 weakly suppresses
  17. L0:1365 weakly excites
  18. L0:257 weakly excites
  19. L0:2899 weakly excites
  20. L0:823 weakly suppresses
  21. L0:1696 weakly excites word endings like 'ed' after a stem
  22. L0:112 weakly suppresses
  23. L0:2456 weakly excites
  24. L0:735 weakly excites

Feeds later layers

  1. L9:2257 weakly excites
  2. L12:1003 weakly suppresses
  3. L27:1552 weakly excites
  4. L6:1317 weakly suppresses "than" after "more to it" comparisons
  5. L27:2457 weakly excites
  6. L7:375 weakly suppresses
  7. L7:2645 weakly excites
  8. L23:1555 weakly excites
  9. L5:2492 weakly excites
  10. L9:1598 weakly suppresses
  11. L11:161 weakly excites
  12. L9:1396 weakly suppresses
  13. L9:2255 weakly suppresses
  14. L8:1455 weakly suppresses
  15. L22:1843 weakly suppresses
  16. L10:2161 weakly excites
  17. L21:2696 weakly excites
  18. L2:2253 weakly suppresses
  19. L3:2892 weakly suppresses
  20. L24:2090 weakly suppresses
  21. L4:1096 weakly excites
  22. L11:918 weakly excites
  23. L4:1133 weakly suppresses
  24. L2:577 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 -2.10 and -2.041 values between -0.45 and -0.391 values between -0.39 and -0.339 values between -0.33 and -0.2714 values between -0.27 and -0.2062 values between -0.20 and -0.14160 values between -0.14 and -0.08773 values between -0.08 and -0.022362 values between -0.02 and 0.04353 values between 0.04 and 0.10111 values between 0.10 and 0.1641 values between 0.16 and 0.2214 values between 0.22 and 0.286 values between 0.28 and 0.35offnegativepositiveoften

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. 讫
  2. ght
  3. ippers
  4. $obj
  5. 利物
  6. 롤
  7. $"
  8. tatto
  9. _taken
  10. oin

Suppresses

  1. afirm
  2. ?=.*
  3. WATCH
  4. orgen
  5. ampus
  6. .isArray
  7. 生活水平
  8. busters
  9. евич
  10. WATCH