Neuron Explorer

Qwen3-0.6B

commas inside number lists

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 lights up on commas that sit inside lists or number sequences, such as "1,2,3" or "143,754". Turning it up made the model get stuck repeating small words, so the steering does not add much.

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

    als are still useful. Arabic numerals share their origin from the numbering system of India.⏎The pattern of Arabic numerals start from 0,1,2,3,4 and it simply goes up in such way. On the other hand

  2. #2 peak

    icata|⏎|Species Authority:||Todd, 1913|⏎|Red List Category & Criteria:||Endangered B1ab(i,ii,iii,v) ver 3.1|⏎|Reviewer(s):||Butchart

  3. #3 peak

    ,182; Aurora, 286,028; Lakewood, 143,754; Fort Collins, 124,665;

  4. #4 peak

    couch or the bed. Shelter,food,clothes,education, and family are all the things that every person deserves. These are all basic needs

  5. #5 peak

    link="none" columns="2" size="medium" ids="2568,2567,2569,2570"]These

  6. #6 peak

    also bought.Page 1 of 1 Start over Page 1 of 1 .Previous page.1/4 x 6 x 6 Stainless Steel Plate,30

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: if an idea/theory//orc////////////////////////////////////////

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

    Fire tube boilers — Biomass boilers; Bubbling fluidized bed boilers; Circulating fluidized bed boilers⏎- Front<|endoftext|>CNY - 中國元,人民幣⏎最常用 CNY 交叉匯率⏎為什麼您

  2. #2 peak

    firms increase revenue.⏎The dynamic pricing version of yield management has the same revenue enhancing goal. As you probably know, all sorts of goods like bicycles, jewelry and detergents are dynamically priced online. One baby clothes vendor changes his prices every

  3. #3 peak

    Studies show influenza viruses can survive on surfaces for up to eight hours, so it is important to use products that are effective against them. Cleaning with soaps or detergents and water physically removes germs, but

  4. #4 peak

    by a variety of health issues among which include genetics, allergies and irritants. Contact dermatitis may be caused by skin irritants such as bath soaps, detergents, laundry soap or bleach. Dermatitis can also be caused by exposure

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. L19:43 strongly excites
  2. L24:1223 strongly excites
  3. L20:1298 strongly suppresses
  4. L2:404 excites
  5. L16:1 excites
  6. L4:384 suppresses
  7. L20:961 suppresses
  8. L20:344 excites hyphen inside compound words
  9. L21:1145 excites
  10. L16:7 suppresses
  11. L19:1826 suppresses
  12. L18:1997 suppresses
  13. L3:395 suppresses
  14. L16:2 suppresses
  15. L2:2051 suppresses
  16. L3:508 suppresses the final zero of decades like '1980s'
  17. L5:97 suppresses
  18. L14:35 suppresses
  19. L0:68 suppresses
  20. L17:118 excites comma after a clause ending
  21. L9:77 excites commas in citations and name lists
  22. L0:290 suppresses
  23. L24:248 excites
  24. L4:3 excites commas after citation fields

Feeds later layers

  1. L27:649 strongly excites commas between numbers in lists
  2. L26:1876 strongly excites
  3. L27:184 suppresses
  4. L26:1478 weakly suppresses
  5. L26:254 weakly suppresses
  6. L27:2439 weakly suppresses slash inside web addresses
  7. L27:1255 weakly excites
  8. L27:405 weakly excites
  9. L27:1793 weakly suppresses
  10. L27:873 weakly suppresses
  11. L26:786 weakly suppresses
  12. L27:1045 weakly suppresses
  13. L27:1463 weakly suppresses
  14. L27:2250 weakly suppresses
  15. L27:986 weakly suppresses
  16. L27:185 weakly excites
  17. L27:1237 weakly excites
  18. L26:2490 weakly excites
  19. L26:711 weakly suppresses 'years' after an age number
  20. L27:721 weakly excites
  21. L27:327 weakly suppresses
  22. L27:647 weakly excites
  23. L27:45 weakly excites
  24. L27:1019 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.

1 values between -5.19 and -4.043 values between -4.04 and -2.895 values between -2.89 and -1.7448 values between -1.74 and -0.583803 values between -0.58 and 0.5719 values between 0.57 and 1.726 values between 1.72 and 2.878 values between 2.87 and 4.022 values between 4.02 and 5.174 values between 5.17 and 6.322 values between 6.32 and 7.472 values between 7.47 and 8.623 values between 10.92 and 12.071 values between 13.23 and 14.381 values between 39.69 and 40.84offnegativepositiveoften

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. 在
  2. 并
  3. 并且
  4. 并在
  5. 则
  6. 并将
  7. 并对
  8. 在一
  9. 在网络
  10. 并通过

Suppresses

  1. naires
  2. ्
  3. ermo
  4. /english
  5. vil
  6. uliar
  7. nowledge
  8. uchos
  9. pants
  10. ্