Neuron Explorer

Qwen3-0.6B

commas in the middle of a sentence

A neuron in layer 19 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 commas that sit in the middle of a sentence and leave the thought open, such as a comma before a list item or a clause that is still going. It is hard to pin down more exactly. Turning it up made the model write long lists joined by commas and "and," and turning it down made it produce lists of letters.

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

    <|endoftext|>The thing I love most about history is its quirky little pieces. Don’t get me wrong- names, dates, and places on a map are all important, and provide a

  2. #2 peak

    see the cervix and the inside of the uterus.⏎- Laparoscopy. This minor procedure uses a laparoscope, a viewing instrument with a light, to see your uterus and pelvis. The laparoscopy is inserted through

  3. #3 peak

    [Insert Poem Title] is no exception. Through her vivid descriptions and thought-provoking themes, Oliver encourages readers to pause, appreciate the beauty of nature, and

  4. #4 peak

    without counting calories and without ramping up your exercise routine,â saysNeal Barnard, M.D., lead author of the study, president of the Physicians Committee, and an adjunct associate professor of medicine at the George Washington University School of Medicine and

  5. #5 peak

    aviolet to be effective, engineers say.⏎That puts those wastewater permit holders in a bind: The law requires them to disinfect, often a costly and complex affair, but their current<|endoftext|>Most of you probably don't remember. I don't,

  6. #6 peak

    in the fume hood). Furthermore, it is best to keep the work space neat and clean in all laboratory operations, particularly those involving the use of hoods, so that any procedure or experiment can be undertaken

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: as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as, as,

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

    make much difference anyway. The best approach is an honest one that doesn’t try to offer a college course in human reproduction in one session. Tell the child that, in a sheltered place inside their bodies, mothers carry a large number of tiny

  2. #2 peak

    typically used to deposit metals or weak insulating materials (such as most nitrides). However, a buildup of insulating material next to the plasma racetrack, and the consequent charging effect of the ionic bombardment, often initiates

  3. #3 peak

    progresses. Among the most important early decisions is whether the needs of the specific clinical trial protocol would be best served by combining all materials in a single kit or if, owing to temperature control and other concerns, it would be more appropriate to divide components

  4. #4 peak

    longer used talc with asbestos. The safety of asbestos-free talc, which is still widely used in many products, is less clear and has been possibly

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. L16:240 strongly excites comma after 'sometimes' or 'the truth'
  2. L14:35 suppresses
  3. L18:1404 excites
  4. L18:1403 excites
  5. L17:118 excites comma after a clause ending
  6. L16:650 suppresses
  7. L18:366 excites comma inside numbers and dates
  8. L15:2506 excites
  9. L15:841 excites
  10. L18:1394 excites
  11. L4:3 excites commas after citation fields
  12. L15:2961 excites
  13. L15:2235 excites
  14. L5:16 excites
  15. L10:335 excites commas in catalog, date, and list metadata
  16. L16:2905 excites
  17. L11:154 suppresses commas after citation numbers
  18. L9:77 excites commas in citations and name lists
  19. L16:1788 suppresses
  20. L17:1405 suppresses
  21. L16:1082 suppresses comma linking two clauses after 'and'
  22. L17:207 suppresses
  23. L15:1159 suppresses ampersand in citation author lists
  24. L15:3010 suppresses

Feeds later layers

  1. L21:344 strongly suppresses
  2. L20:173 strongly suppresses
  3. L27:86 strongly excites
  4. L24:6 strongly excites
  5. L26:236 strongly suppresses
  6. L20:276 strongly suppresses
  7. L26:22 strongly excites the 'res' fragment inside words
  8. L27:286 strongly suppresses
  9. L20:82 strongly excites
  10. L24:642 strongly suppresses
  11. L24:572 strongly suppresses
  12. L27:575 strongly suppresses
  13. L25:438 strongly suppresses
  14. L24:304 suppresses
  15. L26:491 suppresses
  16. L24:216 suppresses
  17. L23:49 suppresses
  18. L25:381 suppresses
  19. L25:328 suppresses
  20. L26:2123 suppresses
  21. L25:969 excites
  22. L27:191 suppresses
  23. L24:468 suppresses
  24. L20:2891 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.

1 values between -25.34 and -24.361 values between -24.36 and -23.381 values between -19.46 and -18.481 values between -18.48 and -17.492 values between -13.57 and -12.592 values between -12.59 and -11.611 values between -11.61 and -10.633 values between -9.64 and -8.661 values between -8.66 and -7.681 values between -7.68 and -6.702 values between -6.70 and -5.723 values between -5.72 and -4.747 values between -4.74 and -3.766 values between -3.76 and -2.7841 values between -2.78 and -1.79116 values between -1.79 and -0.812998 values between -0.81 and 0.17613 values between 0.17 and 1.1543 values between 1.15 and 2.1319 values between 2.13 and 3.1111 values between 3.11 and 4.0910 values between 4.09 and 5.086 values between 5.08 and 6.0610 values between 6.06 and 7.042 values between 7.04 and 8.023 values between 8.02 and 9.001 values between 9.00 and 9.982 values between 10.96 and 11.941 values between 12.93 and 13.91offnegativepositiveoften

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. 而出
  3. eph
  4. ']/
  5. "]
  6. "]/
  7. "/
  8. }
  9. ”
  10. ∷

Suppresses

  1. (),
  2. [],
  3. ++,
  4. **,
  5. ?,
  6. %,
  7. (),
  8. !,
  9. ``,
  10. ~,