Neuron Explorer

Qwen3-0.6B

comma after an introductory phrase

A neuron in layer 12 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 comes after a short lead-in such as a year, a phrase like "as the name may suggest", or "As luck would have it". It seems to mark the point where a sentence is about to continue with its main statement. Steering did not give a clear theme.

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

    2020, shows that even before the COVID-19 pandemic, majorities of citizens reported

  2. #2 peak

    2021, Fox News report, Tucker Carlson asked the question no one is really allowed to ask

  3. #3 peak

    2014, Whales as marine ecosystem engineers, Front Ecol Environ, http://www

  4. #4 peak

    in operational or industrial processes), association mining (customer analytics), and latent variable models (factor analysis).⏎Semi-supervised machine learning⏎As the name may suggest, this model combines elements of supervised and unsupervised learning. It works by using

  5. #5 peak

    on base called in a local beekeeper to clear the swarm, who also cleared up the air on how this happened in the first place. As luck would have it, this particular beekeeper also happens to be a veteran of the U.S.

  6. #6 peak

    the CKSci Curriculum. Modern educational research has provided educators with new directions in communicating the facts and practices of science. Following the release of the Next Generation Science Standards, the Core Knowledge Foundation updated and enhanced the Grades 3–5 science portion of

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 I can't see it, but I know, and I know, and I know, and I know, and I think, and I know, and I know, and I know, and I know, and I think, 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

    outcome of $400,000 compared to an average teacher.⏎Effect of Good Teachers on the Economy⏎What would happen, Hanushek wonders, to aggregate economic growth if poor teachers were replaced with good

  2. #2 peak

    + d / . We might naturally conclude that the total luminosity, the sum of all the individual energies, would be reduced by the same factor. The conclusion, in fact, is valid in the particular

  3. #3 peak

    a disposable one or do you get one that you can recharge? For that matter, why can’t you recharge all of them? What’s the difference? That, turns out, is a matter of science and cost.⏎How Batteries Work⏎

  4. #4 peak

    0 more.⏎Finally, take two of those tiny fragments, each one thousandth of a paper clip or dollar bill, and hold them in your hand. That, according to the Drug Enforcement Administration, is

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. L11:154 suppresses commas after citation numbers
  2. L4:3 excites commas after citation fields
  3. L7:288 excites
  4. L9:77 excites commas in citations and name lists
  5. L5:16 excites
  6. L10:335 weakly excites commas in catalog, date, and list metadata
  7. L11:223 weakly suppresses
  8. L1:2801 weakly excites
  9. L9:1579 weakly suppresses
  10. L4:92 weakly excites
  11. L2:66 weakly excites
  12. L10:1586 weakly suppresses
  13. L9:1900 weakly excites
  14. L10:586 weakly suppresses 'important' and degree words after 'how'
  15. L1:2677 weakly suppresses
  16. L10:422 weakly excites
  17. L11:840 weakly excites
  18. L1:947 weakly excites
  19. L1:1464 weakly excites
  20. L9:749 weakly suppresses
  21. L9:92 weakly excites "and" or "or" joining two items
  22. L5:158 weakly suppresses
  23. L4:21 weakly excites
  24. L6:327 weakly suppresses

Feeds later layers

  1. L13:958 suppresses
  2. L14:2402 suppresses
  3. L22:144 suppresses
  4. L14:1237 suppresses
  5. L14:35 suppresses
  6. L16:846 suppresses comma opening an aside or explanation
  7. L18:366 suppresses comma inside numbers and dates
  8. L13:685 excites
  9. L20:476 suppresses commas that lead into a continuation
  10. L15:1551 excites
  11. L16:2012 suppresses
  12. L14:1060 excites
  13. L15:3049 suppresses
  14. L16:797 weakly suppresses
  15. L16:331 weakly suppresses
  16. L18:1394 weakly suppresses
  17. L21:700 weakly suppresses comma before a new clause
  18. L16:749 weakly excites
  19. L20:1466 weakly suppresses
  20. L15:693 weakly suppresses commas inside numbers and asides
  21. L24:902 weakly suppresses
  22. L23:184 weakly suppresses
  23. L13:114 weakly suppresses
  24. L13:652 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 -7.71 and -7.461 values between -6.19 and -5.941 values between -4.92 and -4.672 values between -4.42 and -4.161 values between -4.16 and -3.911 values between -3.91 and -3.662 values between -3.66 and -3.402 values between -3.40 and -3.153 values between -3.15 and -2.901 values between -2.90 and -2.652 values between -2.65 and -2.395 values between -2.39 and -2.144 values between -2.14 and -1.891 values between -1.89 and -1.635 values between -1.63 and -1.388 values between -1.38 and -1.138 values between -1.13 and -0.8721 values between -0.87 and -0.6253 values between -0.62 and -0.37547 values between -0.37 and -0.112393 values between -0.11 and 0.14777 values between 0.14 and 0.4052 values between 0.40 and 0.658 values between 0.65 and 0.905 values between 0.90 and 1.161 values between 1.16 and 1.412 values between 1.41 and 1.661 values between 2.17 and 2.42offnegativepositiveoften

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. 可以把
  2. 就把
  3. 要把
  4. 取り
  5. 勁
  6. 素
  7. otypes
  8. 巾
  9. 順
  10. ocaust

Suppresses

  1. edia
  2. Es
  3. es
  4. Ed
  5. Es
  6. Ed
  7. ponsored
  8. #=
  9. lication
  10. sert