Neuron Explorer

Qwen3-0.6B

comma before a list continues

A neuron in layer 17 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 fires on a comma that sits in the middle of a list of details, where more items are still to come. Turning it up made the model repeat phrases like 'the house' or 'the news', which does not clearly show the neuron's role.

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

    institutions or research, enabling the transfer of knowledge and technology to socio-economic environment.⏎Project outputs: Awareness of the population (target groups) about the importance of protection, conservation and enhancement<|endoftext|>SoCalGas® is consistently looking for opportunities to reduce our

  2. #2 peak

    for cultural competence and related terms, a brief overview of health disparities and challenges to the common morality, and a discussion of models and frameworks that describe pathways to cultural competence and proficiency.

  3. #3 peak

    into three categories based on the size of the body of water sampled. Reservoirs are defined as bodies of water larger than 1,200 acres, lakes are defined as bodies of water from 10 to 1,2

  4. #4 peak

    may find it significantly more difficult than you feel. But we experience maps on a daily basis. The media utilizes these to pinpoint the position of the latest worldwide turmoil, numerous books include them as drawings, and we consult maps to assist us browse through

  5. #5 peak

    airplanes literally falling from the sky; vehicles could stop functioning, and water, sewer, and electrical networks could all fail at once. The AL

  6. #6 peak

    fiber to the diet has been shown to lower serum cholesterol, decrease the cramps and pain associated with irritable bowel syndrome (or so-called spastic colitis), prevent the development of diverticulosis and diverticulitis

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: and for a and b and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and

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

    -born Hayes followed a fairly typical trajectory to the White House, checking off many of the most common boxes for U.S. presidents—he was educated at Harvard Law School and opened a law practice before serving in the military, then became a Congressman, and

  2. #2 peak

    features of Pyrrhonian Scepticism as Sextus describes it within the Outlines which are of precise philosophical importance: its dedication to the quest for fact and to convinced ideas of rationality, its scope, and its effects for motion and

  3. #3 peak

    , religion, the right to hold public office and the right to vote. Also a place where government is by the people and for

  4. #4 peak

    Begin by studying this choropleth map of Sub-Saharan Africa. Each color corresponds to a World Health Rating determined by the following demographic factors: life expectancy for men and women, infant mortality, crude death rate, people per physician, percentage of 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. L16:850 suppresses
  2. L16:2279 suppresses
  3. L15:412 excites
  4. L16:2733 excites
  5. L16:2247 suppresses
  6. L15:132 suppresses
  7. L12:1756 excites
  8. L15:2277 suppresses
  9. L15:2999 excites
  10. L16:1646 suppresses
  11. L16:2384 suppresses
  12. L16:473 excites
  13. L15:389 suppresses
  14. L16:2200 suppresses
  15. L14:449 weakly suppresses
  16. L11:567 weakly excites
  17. L16:3060 weakly suppresses
  18. L16:2900 weakly suppresses
  19. L15:2009 weakly excites
  20. L16:826 weakly excites
  21. L11:244 weakly excites
  22. L16:1141 weakly excites
  23. L14:447 weakly excites
  24. L9:2337 weakly excites

Feeds later layers

  1. L18:1268 excites
  2. L18:1318 excites
  3. L18:1395 excites
  4. L19:1506 excites
  5. L19:1769 excites
  6. L19:909 excites
  7. L18:2862 excites
  8. L19:867 excites
  9. L19:1082 suppresses
  10. L19:2103 excites
  11. L19:1783 excites
  12. L19:954 weakly excites
  13. L19:766 weakly excites
  14. L20:365 weakly excites
  15. L21:359 weakly excites
  16. L20:2366 weakly excites
  17. L19:1245 weakly excites comma opening a list of parallel items
  18. L19:164 weakly suppresses
  19. L18:851 weakly excites
  20. L21:750 weakly excites
  21. L19:628 weakly suppresses
  22. L20:348 weakly excites
  23. L19:1130 weakly excites
  24. L18:509 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 -5.70 and -5.321 values between -3.79 and -3.411 values between -3.02 and -2.641 values between -2.26 and -1.8810 values between -1.88 and -1.4915 values between -1.49 and -1.1142 values between -1.11 and -0.73151 values between -0.73 and -0.341763 values between -0.34 and 0.041713 values between 0.04 and 0.42114 values between 0.42 and 0.8136 values between 0.81 and 1.1920 values between 1.19 and 1.578 values between 1.57 and 1.9511 values between 1.95 and 2.343 values between 2.34 and 2.723 values between 2.72 and 3.104 values between 3.10 and 3.483 values between 3.48 and 3.874 values between 3.87 and 4.253 values between 4.25 and 4.631 values between 9.23 and 9.61offnegativepositiveoften

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. #error
  2. 🎓
  3. inue
  4. ards
  5. jem
  6. rine
  7. 愆
  8. erval
  9. rif
  10. #[

Suppresses

  1. itu
  2. ive
  3. qued
  4. ppe
  5. حة
  6. ительно
  7. .getCurrentUser
  8. 迦
  9. ã
  10. 动