Neuron Explorer

Qwen3-0.6B

comma after a sentence's subject

A neuron in layer 20 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

Fires on a comma that comes right after the subject of a sentence, just before the writer adds an extra clause, such as 'But when Columba, with the early morning...'. When it was turned up, the model's text fell into a loop of 'and' and 'the'.

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

    gods. But when, by their oppressions, they had stirred up the Egyptians, at length, to revolt, and were driven out of the country, Set, though before an honored deity, was now associated with all that was evil, and

  2. #2 peak

    surcharged with the inspiration of heroism. So the bow was the natural weapon of the simple, perfect physical manhood represented in the idea of Apollo, who, with drawn bow, was the symbol of such manhood displayed in its highest powers

  3. #3 peak

    start deteriorating within the camp, Shorty’s father decides to lead other Japanese Americans in building a baseball field at the camp and forming baseball teams. Shorty, who wasn’t a good baseball player back home, reaches a critical moment and must

  4. #4 peak

    little isle of Oronsay, close by Islay; and here for a moment they thought their future abode was to be. But when Columba, with the early morning, ascending the highest ground on the island, to take what

  5. #5 peak

    ’s modern economic development, the Korean War (1950-1953) allowed everyone in Korea to be on equal footing.⏎Furthermore, Korea, with the birth of a new democratic government, carried out two institutional reforms: land

  6. #6 peak

    . The researchers found that Adidas seemed to have been proactive early, establishing that they wanted to be a leader in the corporate sustainability movement. The researchers determined that Adidas, being an international and European brand, was motivated to become more sustainable due to opportunity

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: I get the case where we have a single case where the two variables are both given as their natural variance case is considered the case where the variance of the is the variance of the is the variance of the case is the variance of

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

    an interaction of nature, nurture and the decisions and choices that people make. Biological programming is your genetics. Genes provide the instructions for structuring amino acids which in turn influence the proteins within our body cells thus programming chemical basis of our biology.

  2. #2 peak

    . The trees that die as a result of rising water levels attract insects, which in turn feed woodpeckers, whose holes later provide homes for other wildlife.⏎

  3. #3 peak

    the group changed its name to the STEMG.⏎The STEMG reports to the ASD Product Support Specification Group (PSSG) which in turn reports to the ASD Services Commission (SVC).<|endoftext|>Biology exquis

  4. #4 peak

    that some writers have focused on one aspect while ignoring the other, while some researchers view these concepts as serving the same purpose. Barber (1998), for example, focused exclusively on civic responsibility and avoided the area of morality. However,

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. L18:2687 suppresses
  2. L19:94 suppresses
  3. L18:3071 excites
  4. L17:2787 excites comma after 'that' or a verb before an aside
  5. L19:2919 excites
  6. L18:174 excites
  7. L18:1864 suppresses
  8. L19:2763 excites
  9. L19:934 excites
  10. L17:1668 excites
  11. L19:993 suppresses
  12. L19:2683 weakly suppresses
  13. L15:883 weakly suppresses
  14. L17:1282 weakly excites
  15. L15:1904 weakly suppresses
  16. L15:621 weakly suppresses
  17. L16:2905 weakly excites
  18. L19:2999 weakly excites
  19. L19:1018 weakly excites
  20. L19:561 weakly excites
  21. L17:116 weakly suppresses
  22. L8:1435 weakly excites
  23. L18:612 weakly excites
  24. L17:1044 weakly excites

Feeds later layers

  1. L26:236 strongly excites
  2. L27:86 strongly suppresses
  3. L26:22 strongly suppresses the 'res' fragment inside words
  4. L27:286 strongly excites
  5. L25:438 strongly excites
  6. L24:572 strongly excites
  7. L24:216 strongly excites
  8. L21:344 strongly excites
  9. L21:393 suppresses unclear: relative 'that' and list items
  10. L25:969 suppresses
  11. L27:191 excites
  12. L27:358 excites
  13. L25:328 excites
  14. L22:368 excites
  15. L26:80 suppresses
  16. L22:940 excites
  17. L25:1221 excites
  18. L26:307 suppresses
  19. L24:304 excites
  20. L21:700 suppresses comma before a new clause
  21. L26:2123 excites
  22. L24:6 suppresses
  23. L27:130 suppresses
  24. L27:575 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.

3 values between -3.53 and -2.964 values between -2.96 and -2.405 values between -2.40 and -1.849 values between -1.84 and -1.2752 values between -1.27 and -0.711241 values between -0.71 and -0.152247 values between -0.15 and 0.41177 values between 0.41 and 0.9852 values between 0.98 and 1.5424 values between 1.54 and 2.1025 values between 2.10 and 2.6716 values between 2.67 and 3.2310 values between 3.23 and 3.7912 values between 3.79 and 4.355 values between 4.35 and 4.922 values between 4.92 and 5.487 values between 5.48 and 6.042 values between 6.04 and 6.602 values between 6.60 and 7.174 values between 7.17 and 7.731 values between 7.73 and 8.291 values between 8.86 and 9.423 values between 9.42 and 9.981 values between 12.79 and 13.361 values between 13.36 and 13.921 values between 15.05 and 15.611 values between 18.42 and 18.98offnegativepositiveoften

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. ,
  2. ,
  3. ,在
  4. ++,
  5. (),
  6. [],
  7. %,
  8. -,
  9. *,
  10. !,

Suppresses

  1. ):⏎⏎
  2. ")⏎
  3. ")
  4. ).*
  5. "]=$
  6. "]
  7. ):⏎
  8. )s
  9. ']
  10. "),⏎