Neuron Explorer

Qwen3-0.6B

unclear: word-start fragments like 'pl' and 'bl'

A neuron in layer 26 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

It fires on the first piece of a word that has been split into smaller chunks, especially chunks beginning with 'pl' or 'bl' such as 'ploughs', 'plinths' and 'blisters'. The pattern is not very clean, and turning it up mostly made the model's writing collapse into odd repeated words.

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

    ’s not expecting food. For this reason, endurance runners frequently consume mushy, easily digestible sources of fuel packed with carbohydrates. These include gel packs, gel chews, jellybeans made specifically for endurance athletes, baby-food pureed fruit packs

  2. #2 peak

    Fossils⏎The coelacanth is a large fish some 1.5 meters long. Its entire body is covered with scales, reminiscent of armor plating. It belongs to the Osteoichthyes class of bony

  3. #3 peak

    dry and itchy skin. Severe cases of dermatitis can result in skin forming crusts or blisters which ooze fluid. There are different types of dermatitis and these are

  4. #4 peak

    balustrades were ornately molded in pre-cast faux stone and lighting was generously provided from electric globes on eleborate cast-iron poles… large plinths marked the foot of the structure at both ends. The four rusticated

  5. #5 peak

    fact that man cannot make any progress without effort and hard work. For example, one who studies whole-heartedly becomes a scholar. Similarly, a farmer who ploughs the farm, sows seeds in it and waters them, is soon

  6. #6 peak

    wagons.⏎With the start of railway rehabilitation in the 1950's, two underframes from scrapped wagons were utilised to build ballast ploughs. These ballast ploughs were towed by ballast trains

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: soulhood tradition and tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition tradition

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

    reports that influenza has resulted in nine to 45 million illnesses, 140,000 to 810, 000 hospitalizations and 12,000 to 61,00

  2. #2 peak

    In fits and spurts, France has sought to extract itself from la Francafrique and to

  3. #3 peak

    vessel walls and lower pressure.⏎High blood pressure is a key risk factor for heart disease, contributing to 47% of heart attacks and 54% of strokes worldwide, the authors write. They note that the risk of heart attack and

  4. #4 peak

    -breaking 40-year-long continuous mission.⏎The Landsat mission is the longest continuous effort to image the Earth, and it has endured many setbacks, failures and near cancellations to get to this moment. Landsat 8 will both rescue the

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. L23:146 excites
  2. L25:986 suppresses
  3. L24:277 excites
  4. L25:425 excites
  5. L22:1542 excites
  6. L24:488 suppresses
  7. L24:368 suppresses
  8. L25:320 excites
  9. L24:268 excites
  10. L19:2956 suppresses
  11. L21:338 suppresses
  12. L22:2438 excites
  13. L24:153 excites
  14. L23:1394 suppresses
  15. L24:109 suppresses
  16. L25:108 suppresses
  17. L20:2997 suppresses
  18. L17:2744 suppresses 'of' after 'those' or 'these'
  19. L18:1556 suppresses
  20. L22:204 excites
  21. L18:1433 excites
  22. L23:919 excites
  23. L25:812 excites
  24. L21:913 weakly suppresses

Feeds later layers

  1. L27:477 excites
  2. L27:60 weakly suppresses
  3. L27:250 weakly excites
  4. L27:533 weakly suppresses
  5. L27:551 weakly suppresses
  6. L27:307 weakly excites
  7. L27:421 weakly suppresses
  8. L27:335 weakly suppresses
  9. L27:729 weakly excites
  10. L27:831 weakly excites
  11. L27:914 weakly excites
  12. L27:480 weakly excites
  13. L27:814 weakly suppresses
  14. L27:829 weakly excites
  15. L27:1902 weakly suppresses
  16. L27:401 weakly excites
  17. L27:687 weakly suppresses
  18. L27:861 weakly excites
  19. L27:546 weakly excites
  20. L27:40 weakly suppresses
  21. L27:1035 weakly excites
  22. L27:521 weakly suppresses
  23. L27:403 weakly excites
  24. L27:930 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 -26.98 and -25.891 values between -22.62 and -21.532 values between -12.79 and -11.702 values between -11.70 and -10.613 values between -10.61 and -9.522 values between -9.52 and -8.424 values between -8.42 and -7.336 values between -7.33 and -6.249 values between -6.24 and -5.1512 values between -5.15 and -4.0633 values between -4.06 and -2.9757 values between -2.97 and -1.87164 values between -1.87 and -0.782251 values between -0.78 and 0.31800 values between 0.31 and 1.40235 values between 1.40 and 2.49119 values between 2.49 and 3.5977 values between 3.59 and 4.6842 values between 4.68 and 5.7719 values between 5.77 and 6.8625 values between 6.86 and 7.9512 values between 7.95 and 9.0510 values between 9.05 and 10.1410 values between 10.14 and 11.233 values between 11.23 and 12.323 values between 12.32 and 13.411 values between 13.41 and 14.503 values between 14.50 and 15.602 values between 15.60 and 16.69offnegativepositiveoften

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. buttons
  2. ifacts
  3. →buttons
  4. rients
  5. Buttons
  6. .buttons
  7. (attributes
  8. buttons
  9. _buttons
  10. constants

Suppresses

  1. ku
  2. 试验区
  3. .cloudflare
  4. 校区
  5. ypo
  6. 旗
  7. 职业
  8. 隅
  9. erk
  10. OPTION