Neuron Explorer

Qwen3-0.6B

comma after who, when or where

A neuron in layer 9 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 wakes up on the comma that comes right after a word like "who", "when" or "where", just before the sentence carries on with its main point. Turning it up or down made the story wander in different directions, and the steered text did not clearly show the same theme.

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

    trees and playing duck-apple at Halloween; burning the clavie (tar barrel) at new year or 'Hallooing Largesse' (where, in

  2. #2 peak

    11 was a victory for the U.S. in the Cold War Space Race with the Soviet Union who, in 1961, had sent Yuri Gagarin to the Cosmos

  3. #3 peak

    cards over three rounds, each new round begins after seven cards have been won. This is accomplished by translating cards in both languages. The winner is the player who, in the final round, translates all of the played cards error-free.⏎Novices

  4. #4 peak

    on a hill overlooking the coast, in order to pay tribute to his Metropolitan, a man who would in the history of Montenegro be identified by the name – St. Peter of C

  5. #5 peak

    per week;⏎- Week 6: Summary assessment.⏎The reason for this approach is to engage your child in the learning process, boost confidence, and show how, in a short period of time, focused work can really pay off!

  6. #6 peak

    Other symptoms include feeling bloated, uncomfortable, and sluggish. In most cases, constipation is temporary and not serious.⏎Many people think they are constipated when, in fact, their bowel movements are regular. For example, some people believe they

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 can see the pattern in the numbers 1 to 12 (which are the numbers 1 to 12 in the sequence) is the part of the pattern itself, which is the result of the multiplication of the numbers

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

    There is some disagreement on the definition of lexicology, as distinct from lexicography.

  2. #2 peak

    about depression and most people don't realize that it is a serious medical issue. Feelings of hopelessness, despair, anxiety, and inappropriate guilt and shame can and do contribute to substance abuse, overeating, oversleeping, and sometimes suicidal

  3. #3 peak

    the receiving end, we re-hash the received data and check that the computed hash equals that sent; if any of the data has changed then (with overwhelming probability), the computed hash value will no longer match the original. Another case is where we

  4. #4 peak

    an integral part of an air handler or it can be an added component to a furnace.⏎The two components are connected with copper refrigeration lines, which basically, transfer heat to the outside, via freon.⏎The efficiency of an air conditioner

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. L8:475 excites
  2. L8:2822 excites
  3. L7:84 suppresses
  4. L5:1871 excites
  5. L7:2538 suppresses
  6. L7:2009 excites
  7. L7:1374 suppresses
  8. L8:511 excites
  9. L7:1098 excites
  10. L7:890 excites
  11. L7:2533 excites
  12. L5:259 suppresses
  13. L1:223 excites unclear: 'selected' and 'shape' in mixed text
  14. L7:499 suppresses
  15. L4:357 excites
  16. L8:1625 suppresses
  17. L8:573 excites
  18. L7:2570 weakly suppresses
  19. L8:678 weakly suppresses
  20. L8:467 weakly suppresses
  21. L5:2625 weakly excites
  22. L8:1610 weakly suppresses
  23. L0:878 weakly suppresses 'because' after a claim
  24. L4:596 weakly excites

Feeds later layers

  1. L10:582 excites
  2. L11:398 suppresses
  3. L10:470 excites
  4. L13:921 excites
  5. L10:100 excites
  6. L10:422 suppresses
  7. L11:948 excites
  8. L11:459 excites
  9. L13:645 excites
  10. L13:78 suppresses
  11. L16:426 suppresses
  12. L10:766 excites
  13. L14:799 excites
  14. L12:1022 weakly suppresses
  15. L17:3068 weakly excites comma after a sentence subject
  16. L10:1449 weakly excites
  17. L14:919 weakly excites
  18. L14:2610 weakly excites
  19. L13:2875 weakly excites
  20. L17:987 weakly excites
  21. L13:2929 weakly excites
  22. L11:2492 weakly excites
  23. L12:1864 weakly excites
  24. L11:1947 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.

2 values between -1.47 and -1.335 values between -1.19 and -1.052 values between -1.05 and -0.925 values between -0.92 and -0.784 values between -0.78 and -0.6424 values between -0.64 and -0.5083 values between -0.50 and -0.36251 values between -0.36 and -0.22617 values between -0.22 and -0.092049 values between -0.09 and 0.05655 values between 0.05 and 0.19121 values between 0.19 and 0.3328 values between 0.33 and 0.4718 values between 0.47 and 0.607 values between 0.60 and 0.7410 values between 0.74 and 0.887 values between 0.88 and 1.024 values between 1.02 and 1.166 values between 1.16 and 1.292 values between 1.29 and 1.431 values between 1.57 and 1.713 values between 1.71 and 1.851 values between 1.85 and 1.981 values between 2.54 and 2.671 values between 2.95 and 3.091 values between 3.92 and 4.06offnegativepositiveoften

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. -posts
  3. _average
  4. -books
  5. //{⏎
  6. }-{
  7. ivre
  8. -facebook
  9. (xi
  10. enquiry

Suppresses

  1. type
  2. amac
  3. TYPE
  4. 类型
  5. IZER
  6. 范
  7. itone
  8. ean
  9. bout
  10. ez