Neuron Explorer

Qwen3-0.6B

comma after a cause or effect clause

A neuron in layer 25 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 sentence describing a cause, such as "too much sodium causes your body to retain water," where the sentence is about to add what it leads to. Turning it up made the model repeat the word "making", which fits that it expects a phrase like "making" or "leading" to follow.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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

    that is the cause of the health concerns.* Sodium is essential in the body to help regulate fluids and blood pressure. Too much sodium causes your body to retain water, which puts extra strain on your heart and blood vessels. By lowering the amount of

  2. #2 peak

    . Eventually, however, it may fail to keep up, resulting in type 2 diabetes.⏎Your genes, excess body fat and lack of exercise cause insulin resistance, but eating the wrong kinds of carbohydrates seems to worsen the problem. In one

  3. #3 peak

    . By performing weight-training exercises, you can build strength in your muscles, increase the amount of lean muscle in your body, manage your weight more easily and

  4. #4 peak

    . However, if you’re able to use<|endoftext|>Need an excuse to treat yourself to dessert? A daily dose of chocolate or cocoa powder could help lower blood pressure, say the authors of a new systematic review by the Cochrane Library.⏎

  5. #5 peak

    agricultural practices wherein the soil is consistently turned over. By doing so, each layer of the soil gets exposed to the sun, and leads to soil dehydration. Soil comprises various bacteria, fungi and worms that are

  6. #6 peak

    are never filled. Such shortcomings in medication adherence can decrease the effectiveness of a treatment. This increases the likelihood that a condition will worsen, even to the point of death.⏎But these figures fail to consider the numerous challenges

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 always associate the numbers 7 and 8 with the numbers 8 and 9 always again, these numbers always hold these same numbers even if these numbers are also written in a different format even if these numbers also can be written in

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

    , but in one’s chest, one’s hands, and one’s feet, all together.”⏎Long describes Homer’s overall style to be that of ‘sensuous immediacy’, writing: ‘Homer delights in the vibrancy of every

  2. #2 peak

    am is introduced to Frances Henri, a lady teacher who instructs the girls in sewing, as his new student. Sensible, industrious and desolate, this girl slowly gains his admiration and Mad

  3. #3 peak

    He became a billionaire by founding the Standard Oil Company, which was the largest oil company in the world at the time. Rockefeller was also a great philanthropist, donating millions of

  4. #4 peak

    in Washington, D.C., and became a house painter, enjoying a reasonably comfortable life. Turnage had a more hardscrabble existence in New York City, where he worked as a watchman, waiter, and other occupations.⏎Their narratives

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. L19:1036 suppresses
  2. L18:366 excites comma inside numbers and dates
  3. L22:1425 suppresses
  4. L14:35 suppresses
  5. L22:144 suppresses
  6. L17:118 excites comma after a clause ending
  7. L20:476 excites commas that lead into a continuation
  8. L18:892 suppresses
  9. L22:1414 suppresses
  10. L16:846 excites comma opening an aside or explanation
  11. L19:726 suppresses comma inside large numbers
  12. L19:944 suppresses
  13. L16:240 excites comma after 'sometimes' or 'the truth'
  14. L18:1688 excites
  15. L22:940 suppresses
  16. L21:200 suppresses
  17. L19:291 weakly suppresses 'and' joining two full clauses
  18. L19:1380 weakly excites
  19. L19:1361 weakly suppresses
  20. L19:2175 weakly suppresses
  21. L17:3021 weakly suppresses
  22. L19:94 weakly excites
  23. L19:383 weakly suppresses commas after a clause or list item
  24. L20:1832 weakly suppresses 'as' introducing a reason

Feeds later layers

  1. L27:410 excites
  2. L27:677 excites
  3. L26:686 excites
  4. L26:252 excites
  5. L27:1890 excites
  6. L26:1304 weakly excites
  7. L27:1545 weakly excites
  8. L26:708 weakly excites
  9. L26:1296 weakly suppresses word-start 'mult' fragment
  10. L26:979 weakly suppresses
  11. L26:2677 weakly excites
  12. L27:1148 weakly excites
  13. L27:566 weakly excites
  14. L27:409 weakly suppresses
  15. L26:432 weakly excites
  16. L26:2383 weakly suppresses
  17. L27:2982 weakly excites
  18. L26:2881 weakly suppresses
  19. L27:248 weakly suppresses
  20. L27:1112 weakly excites
  21. L26:2104 weakly suppresses
  22. L26:764 weakly suppresses
  23. L26:1444 weakly suppresses
  24. L27:1515 weakly suppresses

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 -23.20 and -22.021 values between -20.83 and -19.652 values between -17.27 and -16.091 values between -16.09 and -14.901 values between -14.90 and -13.722 values between -13.72 and -12.532 values between -12.53 and -11.343 values between -11.34 and -10.164 values between -10.16 and -8.977 values between -8.97 and -7.796 values between -7.79 and -6.6011 values between -6.60 and -5.4112 values between -5.41 and -4.2324 values between -4.23 and -3.0453 values between -3.04 and -1.86276 values between -1.86 and -0.672515 values between -0.67 and 0.52466 values between 0.52 and 1.70200 values between 1.70 and 2.89122 values between 2.89 and 4.0766 values between 4.07 and 5.2653 values between 5.26 and 6.4521 values between 6.45 and 7.6322 values between 7.63 and 8.8210 values between 8.82 and 10.007 values between 10.00 and 11.195 values between 11.19 and 12.385 values between 12.38 and 13.562 values between 13.56 and 14.751 values between 15.93 and 17.121 values between 17.12 and 18.311 values between 18.31 and 19.491 values between 20.68 and 21.862 values between 23.05 and 24.23offnegativepositiveoften

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. making
  2. making
  3. giving
  4. suggesting
  5. leading
  6. allowing
  7. implying
  8. leaving
  9. indicating
  10. confirming

Suppresses

  1. 体
  2. hari
  3. APTER
  4. 世
  5. ownt
  6. stands
  7. 就好
  8. 쯤
  9. ote
  10. estroy