Neuron Explorer

Qwen3-0.6B

colon after a metadata field name

A neuron in layer 7 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

The neuron fires on the colon that comes right after a label in a structured record, such as 'Authors:', 'Name:' or 'Family:'. The steered text didn't pick up anything related, so the effect is only visible in the examples.

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

    the same would be achieved with the following code:⏎TYPE List = REF RECORD next: List; value: INTEGER; END;⏎VAR x := NEW(List, value:=1); BEGIN x.next :=

  2. #2 peak

    Army. Indiana Infantry Regiment, 84th (1862-1865) | Purdue University Libraries, Archives and Special Collections⏎Name: United States.

  3. #3 peak

    Prev Med Rep⏎Authors: Altman EA, Lee KL, Hecht CA, Hampton KE, Moreno G

  4. #4 peak

    J. (Robert John), 1801-1869⏎Articles of Agreement and Specifications, Topeka School House Construction⏎Authors: New England Emigrant Aid Company⏎Date: June 9, 18

  5. #5 peak

    factor receptor)|⏎Crystallographic structure of MET. PDB rendering based on 1r0p.⏎|External IDs||ChEMBL: GeneCards:|⏎|RNA expression pattern|⏎c-Met (MET or MNNG HOS

  6. #6 peak

    opoda (bipedal predatory dinosaurs) >⏎Coelurosauria > Maniraptora > Aves⏎Order: Passeriformes > Family: Paridae⏎Southern black tit male, Thornybush Game Reserve,⏎South

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 have to be on the point of the point. I have to be on the point of the point. I have to be on the point of the point. I have to be on the point of the point again. I have to

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

    6 Not Acceptable response because only German and English versions of the document are available.⏎GET /news HTTP/1.1 Host: www.example.ai Accept-Language: fr⏎HTTP/1.1 <|endoftext|>The move to electronic health records is

  2. #2 peak

    sinuous features suggestive of downslope flow, also potentially caused by creep of ground ice.⏎Image information: VIS instrument. Latitude -40.4, Longitude 132.5 East (227.5 West). 3

  3. #3 peak

    14 June 2006.⏎Link: http://www.un.org/apps/news/story.asp?NewsID=18850&Cr=tsunami&Cr

  4. #4 peak

    learn more about Caddo history and culture, https://scholarworks.sfasu.edu/cgi/viewcontent.cgi?article=4012&context=ita⏎Visit the THC’s Learning Resource page for home-related lesson plans and activities

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. L6:74 excites
  2. L3:77 excites
  3. L6:956 weakly suppresses
  4. L6:314 weakly excites
  5. L6:984 weakly suppresses
  6. L6:436 weakly excites
  7. L4:402 weakly suppresses
  8. L5:683 weakly suppresses
  9. L1:1346 weakly excites
  10. L5:1558 weakly suppresses
  11. L6:65 weakly excites
  12. L3:2783 weakly excites
  13. L1:174 weakly excites
  14. L4:193 weakly excites
  15. L5:585 weakly excites
  16. L3:2543 weakly suppresses
  17. L5:284 weakly excites
  18. L2:571 weakly suppresses
  19. L2:243 weakly excites
  20. L3:2182 weakly excites
  21. L5:2467 weakly suppresses
  22. L1:2626 weakly suppresses
  23. L6:2572 weakly excites
  24. L1:2791 weakly suppresses

Feeds later layers

  1. L8:382 excites
  2. L14:216 excites
  3. L11:681 weakly excites
  4. L13:449 weakly excites
  5. L13:1147 weakly excites
  6. L24:1363 weakly excites
  7. L24:30 weakly suppresses
  8. L8:781 weakly excites
  9. L13:2987 weakly excites
  10. L10:77 weakly suppresses
  11. L14:1101 weakly suppresses
  12. L15:1661 weakly suppresses
  13. L12:940 weakly excites
  14. L23:3040 weakly suppresses
  15. L17:2364 weakly excites
  16. L8:2691 weakly suppresses
  17. L24:6 weakly suppresses
  18. L15:2422 weakly excites
  19. L21:2675 weakly suppresses
  20. L18:1854 weakly excites
  21. L13:322 weakly suppresses
  22. L10:936 weakly excites
  23. L23:1429 weakly excites
  24. L16:913 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.

4 values between -0.58 and -0.552 values between -0.55 and -0.524 values between -0.52 and -0.492 values between -0.49 and -0.457 values between -0.45 and -0.4210 values between -0.42 and -0.399 values between -0.39 and -0.3625 values between -0.36 and -0.3344 values between -0.33 and -0.3046 values between -0.30 and -0.2692 values between -0.26 and -0.23123 values between -0.23 and -0.20153 values between -0.20 and -0.17188 values between -0.17 and -0.14232 values between -0.14 and -0.10283 values between -0.10 and -0.07350 values between -0.07 and -0.04416 values between -0.04 and -0.01511 values between -0.01 and 0.02382 values between 0.02 and 0.06222 values between 0.06 and 0.09209 values between 0.09 and 0.12172 values between 0.12 and 0.15127 values between 0.15 and 0.1891 values between 0.18 and 0.2166 values between 0.21 and 0.2540 values between 0.25 and 0.2844 values between 0.28 and 0.3125 values between 0.31 and 0.347 values between 0.34 and 0.378 values between 0.37 and 0.415 values between 0.41 and 0.441 values between 0.44 and 0.472 values between 0.47 and 0.501 values between 0.50 and 0.531 values between 0.53 and 0.571 values between 0.57 and 0.601 values between 0.63 and 0.662 values between 0.66 and 0.69offnegativepositiveoften

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. cul
  2. -degree
  3. reet
  4. orn
  5. =en
  6. holm
  7. -than
  8. uo
  9. ackle
  10. Cheers

Suppresses

  1. 指挥部
  2. .ResponseWriter
  3. 坼
  4. 妯
  5. patches
  6. 趋势
  7. KY
  8. 会议精神
  9. urvey
  10. 女王