Neuron Explorer

Qwen3-0.6B

number in days or time spans

A neuron in layer 23 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the number in phrases that say how many days or nights something lasts, like "28 days", "seven days" or "eight nights". When it is turned down the model drifted toward talk about quarters and years instead.

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

    the benefits of managing menstrual hygiene (MHM).⏎Menstrual Hygiene Day was chosen for May 28 because most women’s menstrual cycles last 28 days and their menstruation phase last five days

  2. #2 peak

    will see new skin underneath.⏎Cold sores typically last about 7 to 10 days. It is contagious and can be spread by direct contact with other people.

  3. #3 peak

    better.⏎The fever associated with the flu typically lasts between three and five days, but fatigue and other symptoms can persist for weeks.⏎A sinus infection lasts longer than seven days in most cases. Many patients experience “double worsening,” meaning they begin to

  4. #4 peak

    4 days, depending on the species and the air temperature. In tropical conditions, the first three stages tend to be longer, usually 10 to 14 days.⏎The adult female Anopheles mosquito can live up to a month,

  5. #5 peak

    start creating your Spark projects and collaborating with other data scientists. When you sign up, you get free access to Watson Studio and all other IBM services for 30 days. Start now and take advantage of this offer.⏎- Learn how it performs

  6. #6 peak

    easier to understand and encourages participants to change their behaviour patterns.⏎Set a deadline for planning/writing and rehearsing the role-play, giving participants at least 90 minutes to prepare. A late afternoon or evening presentation saves time and often provides the

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 am a *centennial* (centennial is a term I'm semi-annually and a semi-centennial... acentennial? Acentennial? Acent... acent... acent... *cent*... *

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

    society is trying to help blacks lift themselves up to be able to enjoy a more prosperous and satisfying life. Much progress has been made in this respect in the last half-century but there is clearly still a long way to go in achieving full equality with

  2. #2 peak

    to) balance the macro world of nature and the micro world within each human.”⏎This year’s ceremony was more lavish than the one Rendeh witnessed nearly half a century ago.⏎“That last one was not so large as this. We were poorer

  3. #3 peak

    on owner-cultivator principles and a compulsory education system for school-age children to be enrolled at a primary school, at the very least.⏎In the past half century, Korea’s rapid economic growth has evolved around

  4. #4 peak

    [on sheet with] Plan du Havre de Noel, dans la Terre de Kerguelen, by Cook/Benard, from Atlas du Troisi<|endoftext|>One of

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. L20:1335 suppresses
  2. L22:1954 suppresses
  3. L14:1088 excites
  4. L22:663 suppresses
  5. L20:1355 suppresses
  6. L17:2544 suppresses
  7. L21:1284 excites
  8. L19:1720 suppresses
  9. L20:2942 suppresses
  10. L22:941 suppresses
  11. L21:708 suppresses
  12. L17:2964 suppresses
  13. L22:1998 suppresses
  14. L21:375 suppresses
  15. L20:1133 suppresses
  16. L20:2953 suppresses
  17. L4:2982 excites
  18. L20:1680 suppresses
  19. L21:620 weakly excites
  20. L20:1155 weakly excites
  21. L7:229 weakly excites first digit of a split number
  22. L5:50 weakly excites
  23. L1:2605 weakly excites
  24. L1:2540 weakly suppresses

Feeds later layers

  1. L24:1368 suppresses
  2. L26:1448 suppresses
  3. L25:1717 weakly suppresses
  4. L27:1105 weakly suppresses
  5. L27:2995 weakly suppresses
  6. L24:450 weakly suppresses
  7. L24:2891 weakly suppresses
  8. L26:2733 weakly suppresses
  9. L25:1323 weakly suppresses
  10. L25:3005 weakly excites
  11. L24:2365 weakly excites
  12. L25:1377 weakly excites
  13. L24:2941 weakly suppresses
  14. L27:1449 weakly excites
  15. L26:2331 weakly suppresses
  16. L26:284 weakly suppresses
  17. L25:1013 weakly suppresses parenthesis introducing an acronym
  18. L25:804 weakly suppresses
  19. L24:478 weakly suppresses word fragment inside a longer word
  20. L25:2729 weakly suppresses
  21. L27:566 weakly suppresses
  22. L25:2346 weakly suppresses
  23. L25:2996 weakly suppresses
  24. L24:282 weakly suppresses the 'conc' word fragment

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.

2 values between -34.22 and -32.771 values between -32.77 and -31.331 values between -26.99 and -25.552 values between -25.55 and -24.101 values between -22.66 and -21.211 values between -19.77 and -18.331 values between -18.33 and -16.881 values between -16.88 and -15.442 values between -15.44 and -13.992 values between -13.99 and -12.554 values between -12.55 and -11.104 values between -11.10 and -9.664 values between -9.66 and -8.217 values between -8.21 and -6.775 values between -6.77 and -5.3214 values between -5.32 and -3.8818 values between -3.88 and -2.4342 values between -2.43 and -0.993468 values between -0.99 and 0.46279 values between 0.46 and 1.9012 values between 1.90 and 3.358 values between 3.35 and 4.795 values between 4.79 and 6.245 values between 6.24 and 7.684 values between 7.68 and 9.139 values between 9.13 and 10.571 values between 10.57 and 12.021 values between 12.02 and 13.461 values between 13.46 and 14.911 values between 14.91 and 16.351 values between 17.80 and 19.241 values between 22.13 and 23.58offnegativepositiveoften

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. rokes
  2. ivity
  3. -days
  4. >Data
  5. Hanna
  6. 面孔
  7. 脸
  8. Brains
  9. exhaustion
  10. riff

Suppresses

  1. 月
  2. ⽉
  3. 季度
  4. 月至
  5. 季
  6. 年
  7. year
  8. quarter
  9. month
  10. _month