Neuron Explorer

Qwen3-0.6B

comma inside numbers with thousands

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 the comma that splits thousands in numbers, such as 1,904 ft or 52,000,000 dogs, and it is often in measurements and statistics. Turning it up produced repeated zeros, which is a sign of a broken output rather than a clear 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

    on those of the West Coast and Marlborough.⏎New Zealand’s highest waterfall⏎The country’s highest waterfall, at 580.3 m (1,904 ft), bears the name of the Scot who discovered it in

  2. #2 peak

    give the park its name.⏎Murchison Falls National Park as it is known today, covers an area of 3,860 km2 (1,557 mi2). It is situated in Northwestern Uganda and is

  3. #3 peak

    location within the U.S.⏎|Named for||Weber River|⏎|• Total||659 sq mi (1,710 km2)|⏎|• Land||576 sq mi

  4. #4 peak

    is permanent and your child will need lifelong treatment.⏎The most common causes are⏎These abnormalities are not inherited from the parents.⏎About 1 in every 2,000 to 4,000 babies is born with congenital

  5. #5 peak

    . Top: Location of the Republic of Palau. Bottom: Location of IS39 array.⏎Few of the islands are inhabited and less than 20,000 people are lucky enough to call Palau home. Tourists from

  6. #6 peak

    single event and not an every day occurrence. Unfortunately, dog attacks are very common and happen every day.⏎There are more than 52,000,000 dogs in the United States alone. Approximately one-third of all homes

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: to the first, only to a second, to a third, to a fourth, to a fifth, to a sixth, ... a first, second, a second, second second, second ... ... to a first, second, third,

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

    as mementos of their visit.⏎This crested souvenir cup was sold by Mrs Emily Yalden who owned a shop selling china and glassware at 17, The Square, Winchester between 1887 and 19

  2. #2 peak

    $\quad+\quad$ or $\quad-$<|endoftext|>Design for a Mosaic⏎Albert Joseph Moore, ARWS 1841-1893⏎According to The Studio, this and

  3. #3 peak

    ZI (1948–2009)⏎He was born at Sáromberke / Dumbrăvioara. He was 42 when, in Marosvásárhely / Târg

  4. #4 peak

    the afternoon on March 19, 1935, an employee at the Kress Five and Ten store at 256 W. 125th Street (just across the street from the Apollo Theater) caught

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. L18:1391 excites
  2. L8:20 suppresses
  3. L17:533 suppresses
  4. L17:128 excites
  5. L18:1758 suppresses
  6. L15:2051 suppresses
  7. L14:542 excites
  8. L12:713 suppresses
  9. L14:1829 suppresses
  10. L4:205 excites the digit 2 inside citations
  11. L16:286 suppresses
  12. L18:617 excites
  13. L16:724 excites
  14. L13:33 excites
  15. L12:2528 suppresses digits inside numbers
  16. L15:12 excites
  17. L6:2363 suppresses
  18. L17:167 suppresses
  19. L18:1259 weakly excites
  20. L17:259 weakly excites
  21. L8:465 weakly suppresses
  22. L16:35 weakly excites 'not' in 'but not limited to'
  23. L10:810 weakly suppresses
  24. L10:1120 weakly suppresses

Feeds later layers

  1. L22:332 excites
  2. L26:413 excites
  3. L27:153 suppresses the spot just before a historical number
  4. L27:862 excites
  5. L25:543 suppresses
  6. L24:371 excites
  7. L27:1196 excites
  8. L21:1350 suppresses
  9. L24:453 suppresses
  10. L25:182 excites
  11. L27:1440 excites
  12. L20:2048 suppresses
  13. L23:615 suppresses
  14. L23:510 excites
  15. L27:294 excites
  16. L20:761 excites
  17. L25:212 excites
  18. L27:1153 excites
  19. L25:1738 excites
  20. L23:2945 excites the '2' in '20th' or '21st' century
  21. L25:232 excites
  22. L20:2667 suppresses
  23. L25:92 weakly excites
  24. L23:924 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.

1 values between -12.29 and -11.232 values between -8.07 and -7.012 values between -7.01 and -5.964 values between -5.96 and -4.904 values between -4.90 and -3.856 values between -3.85 and -2.797 values between -2.79 and -1.74480 values between -1.74 and -0.683310 values between -0.68 and 0.3727 values between 0.37 and 1.4315 values between 1.43 and 2.4910 values between 2.49 and 3.5411 values between 3.54 and 4.607 values between 4.60 and 5.657 values between 5.65 and 6.717 values between 6.71 and 7.761 values between 7.76 and 8.822 values between 8.82 and 9.872 values between 10.93 and 11.981 values between 11.98 and 13.041 values between 18.31 and 19.371 values between 28.87 and 29.92offnegativepositiveoften

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. elsen
  2. itzer
  3. 0
  4. onSave
  5. (Parcel
  6. entialAction
  7. addAction
  8. 0
  9. наци
  10. compart

Suppresses

  1. oted
  2. just
  3. ested
  4. ost
  5. op
  6. otes
  7. 另有
  8. usty
  9. .screen
  10. structured