Neuron Explorer

Qwen3-0.6B

capital L starting a word

A neuron in layer 27 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 reacts to a capital L at the start of a word, such as Listeria, Lying or Lizards. Its opposite side reacts to a small l in the middle of words like luster. Turning it up produced repetitive, unrelated text.

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

    risks.⏎Why is drinking raw milk risky?⏎Milk that has not been pasteurized can contain harmful bacteria such as Salmonella, E. coli, and Listeria, which are responsible for causing

  2. #2 peak

    small amounts of clear liquids⏎If you have morning sickness during pregnancy, ask your doctor about the many possible treatments.⏎The following may help treat motion sickness:⏎- Lying down⏎- Over-the-counter antihistamines (such as Dramamine

  3. #3 peak

    their meaningful contexts, finding it in interesting media and so on. Only when love is strong, memory work can follow.⏎Yesterday, we worked with mirror books. L took notes for the story. Meanwhile, I just want to mention, as an

  4. #4 peak

    diseases contained while doing research. These cabinets feature fundamental technologies including HEPA filters, Laminar Airflow, workstations, ventilated cabinets, and exhausts.

  5. #5 peak

    el Bag Nuzest Black Face Mask NUZEST BlenderBottle® SHAKER CUPS Nuzest Gift Card DIGESTION SUPPORT RECIPE from Lilsipper | Digital Download 7 Day Vegan Meal Plan | Digital download⏎

  6. #6 peak

    iusLeadership | Messenger birds | Lugh Long-arm | Odin | Rigveda | Thor | Odin and Tyr |

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: if I take the partition function partition function partition function partition function partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition partition

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

    ance, sparkle, and translucency. Shades and colours are characterized without all the properties such as a gloss, luster and transparency, to examine the identical.⏎People who operate in the textile industry

  2. #2 peak

    be set side by side without gaps. The effect of this unbroken display of diamonds in a channel setting can create a startlingly beautiful light display accent. The luster and clarity, as well as the muted display of fire, that is found

  3. #3 peak

    High grade ores and ores resistant to cyanide leaching at coarse particle⏎Gold processing Britannica.com⏎The mining of alluvial deposits and, later, lode or vein deposits required crushing prior to gold extraction, and this consumed immense amounts

  4. #4 peak

    , porous bone enclosing numerous large spaces that give a honeycombed or spongy appearance. The bone matrix, or framework, is organized into a three-dimensional latticework of bony processes, called trabeculae, arranged along

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. L23:1760 suppresses
  2. L21:252 excites
  3. L24:2164 excites
  4. L18:198 suppresses
  5. L23:206 weakly suppresses
  6. L19:593 weakly excites
  7. L12:1814 weakly suppresses
  8. L22:151 weakly suppresses
  9. L24:105 weakly suppresses
  10. L24:191 weakly suppresses
  11. L23:1201 weakly suppresses
  12. L22:3039 weakly excites
  13. L6:2013 weakly suppresses
  14. L16:1040 weakly suppresses
  15. L6:2277 weakly suppresses
  16. L2:1433 weakly suppresses
  17. L5:2412 weakly excites
  18. L0:167 weakly suppresses
  19. L9:1880 weakly excites
  20. L5:468 weakly excites
  21. L26:858 weakly suppresses
  22. L3:1219 weakly excites
  23. L25:1207 weakly excites
  24. L0:2661 weakly excites

Feeds later layers

None: this is the last layer.

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.

2 values between -24.53 and -22.052 values between -9.64 and -7.1611 values between -7.16 and -4.6812 values between -4.68 and -2.193446 values between -2.19 and 0.29400 values between 0.29 and 2.7712 values between 2.77 and 5.257 values between 5.25 and 7.745 values between 7.74 and 10.225 values between 10.22 and 12.701 values between 12.70 and 15.181 values between 15.18 and 17.661 values between 27.59 and 30.071 values between 37.52 and 40.001 values between 59.86 and 62.341 values between 72.27 and 74.75offnegativepositiveoften

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. зрения
  2. POSITE
  3. Goals
  4. alle
  5. ulated
  6. ald
  7. setSupportActionBar
  8. .Manager
  9. 当前位置
  10. .NewRequest

Suppresses

  1. 량
  2. (:,:,
  3. dela
  4. berg
  5. 准
  6. 高昂
  7. 不开
  8. .ps
  9. /part
  10. леч