Neuron Explorer

Qwen3-0.6B

unclear: first pieces of split words

A neuron in layer 26 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

It mostly lights up on the first chunk of a word that the model has to finish, such as 'par' in 'parmesan' or 'conce' in 'conceives'. It doesn't clearly respond to one topic, and turning it down or up produced repetitive or off-topic text.

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

    ry: Saute broccoli florets with cooked quinoa, garlic, ginger, and soy sauce for a protein-packed and satisfying meal.⏎- Roasted broccoli with parmesan cheese: Toss broccoli flo

  2. #2 peak

    rets with olive oil, garlic, and parmesan cheese and roast in the oven for a crispy and flavorful side dish.⏎

  3. #3 peak

    'Avola has spread to nearby cities (Noto, Rosolini, Pachino, Ispica and Victoria) to affect areas of the province of Caltanissetta and Trapani (1968).⏎Prior to

  4. #4 peak

    , East Java; a green fashion show in Korea; and a project aimed at developing smart grid systems in Vietnam.⏎Young leaders can make a difference⏎Siti Nur Alliah, a British Council Climate Champion from Indonesia, is tackling a burning issue

  5. #5 peak

    of Mines, Golden, Colorado 1979:43 pages⏎Dust control techniques in coal mines were investigated to perfect methods of generating monodisperse coal dust, develop methods for directly measuring supersaturation, make improvements in

  6. #6 peak

    jaw that surround the roots of the teeth.⏎The curving part of the jaw into which the teeth are rooted.⏎The socket in the

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: they were half the room half the room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they

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

    the Peace Corps. He created the Foundation Heart Ventures-Grupo de Ap<|endoftext|>is an elementary subatomic particle with infinitesimal mass⏎(less than

  2. #2 peak

    out night patrols.⏎Ten Nieuport 80s went to Brazil in November 1918, where they were used by the Escola de Aviacao Militar at Campos dos Afonsos. Brazil also received a number

  3. #3 peak

    8, in the Nihon Shoki 日本書紀 (720) 9 and in the Hizen-no-kuni fudoki 肥前國風土記 (first part of the 8th century)

  4. #4 peak

    bu Anne mau makan di KFC.⏎Harimau tidak mau berlari ke rumah.⏎Perempuan muntah pohon.⏎Students

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. L16:454 suppresses
  2. L24:851 excites
  3. L23:389 suppresses
  4. L22:819 suppresses
  5. L18:541 weakly suppresses
  6. L11:2040 weakly suppresses
  7. L21:40 weakly excites
  8. L20:174 weakly excites
  9. L6:328 weakly suppresses
  10. L6:344 weakly suppresses
  11. L21:160 weakly suppresses
  12. L23:352 weakly suppresses
  13. L25:57 weakly excites
  14. L1:1811 weakly suppresses
  15. L21:266 weakly excites
  16. L5:50 weakly excites
  17. L2:2154 weakly suppresses
  18. L19:853 weakly excites
  19. L0:26 weakly excites
  20. L18:102 weakly suppresses
  21. L16:463 weakly suppresses 'for' introducing a purpose or target
  22. L14:95 weakly excites
  23. L16:134 weakly suppresses
  24. L25:362 weakly excites

Feeds later layers

  1. L27:1233 weakly excites
  2. L27:796 weakly excites
  3. L27:1504 weakly excites
  4. L27:2210 weakly suppresses
  5. L27:779 weakly suppresses
  6. L27:783 weakly excites
  7. L27:2604 weakly excites
  8. L27:1909 weakly excites
  9. L27:866 weakly excites 'ch' starting a word fragment
  10. L27:752 weakly suppresses
  11. L27:568 weakly excites
  12. L27:2264 weakly excites
  13. L27:2061 weakly suppresses
  14. L27:1937 weakly excites
  15. L27:842 weakly excites
  16. L27:18 weakly excites
  17. L27:2190 weakly suppresses
  18. L27:1881 weakly suppresses
  19. L27:35 weakly suppresses
  20. L27:2330 weakly excites
  21. L27:2380 weakly suppresses
  22. L27:2 weakly excites
  23. L27:275 weakly suppresses
  24. L27:2874 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 -36.66 and -35.351 values between -14.41 and -13.111 values between -13.11 and -11.804 values between -9.18 and -7.875 values between -7.87 and -6.563 values between -6.56 and -5.264 values between -5.26 and -3.956 values between -3.95 and -2.647 values between -2.64 and -1.331028 values between -1.33 and -0.022808 values between -0.02 and 1.2926 values between 1.29 and 2.603 values between 2.60 and 3.905 values between 3.90 and 5.212 values between 5.21 and 6.521 values between 7.83 and 9.141 values between 9.14 and 10.452 values between 14.37 and 15.68offnegativepositiveoften

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. etta
  2. >window
  3. met
  4. cost
  5. ett
  6. 师
  7. 武
  8. 写字
  9. -cost
  10. 谋

Suppresses

  1. #{@
  2. utes
  3. ute
  4. ươi
  5. itional
  6. =#{
  7. /#{
  8. ց
  9. себе
  10. IDER