Neuron Explorer

Qwen3-0.6B

sentence-final periods vs dots in code

A neuron in layer 2 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 reacts to a full stop that begins a new sentence in ordinary prose, and reacts negatively to dots inside code, file names and web addresses. Turning it up did not produce an obvious theme in the model's writing.

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

    . For example, it’s a good idea to place the loading and unloading bay

  2. #2 peak

    . of the Afro-Asiatic language family that includes Hebrew, Arabic, Ar

  3. #3 peak

    . The hotel also had several amenities, including a casino, a golf course, and

  4. #4 peak

    . If there are no local brands to satisfy a need, then consumers will purchase foreign

  5. #5 peak

    . Election day was largely peaceful and approximately 81% of voters cast ballots,

  6. #6 peak

    . The new method, published today in the journal Nature, makes it possible to store

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 need to get it. This is a problem I have. The problem is to find the solution. The problem is to solve. The problem is to be solved. The problem is to be solved. The problem is to be solved.

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

    issolved output=gadm36_dissolved_nocats option=del⏎Then, add unique categories to them.⏎v.category input=gadm36_dissolved_nocats output=gadm36

  2. #2 peak

    category for all the areas.⏎To address this problem, we just need to recreate categories for all the features. First, delete the single category from them.⏎v.category input=gadm36_d

  3. #3 peak

    analysis of the tooth surface cutting movement, the revolution and rotation of the blade are further analyzed, as shown in the figure.<|endoftext|>Source: The Independent – independent.co.uk⏎Up, Up and Away on Google’s Wi-Fi Balloon⏎By

  4. #4 peak

    Grid: SO281914⏎Mapcode National: GBR B3.GK6K⏎Mapcode Global: VH75W.W0YX

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. L0:36 strongly excites
  2. L0:387 suppresses
  3. L0:2838 suppresses
  4. L1:38 suppresses
  5. L1:138 suppresses
  6. L0:676 weakly excites
  7. L1:1414 weakly excites
  8. L0:1224 weakly suppresses
  9. L0:479 weakly excites
  10. L0:1301 weakly suppresses
  11. L0:87 weakly suppresses
  12. L0:1605 weakly suppresses
  13. L0:178 weakly suppresses
  14. L1:321 weakly suppresses unclear: period at fragment start
  15. L1:2500 weakly excites
  16. L1:1149 weakly suppresses
  17. L0:1844 weakly excites
  18. L0:1192 weakly excites
  19. L1:331 weakly suppresses
  20. L0:2833 weakly excites
  21. L0:350 weakly excites
  22. L1:2973 weakly suppresses
  23. L0:0 weakly excites
  24. L0:1670 weakly excites

Feeds later layers

  1. L8:143 strongly suppresses
  2. L15:313 strongly suppresses decimal point inside numbers
  3. L4:7 strongly suppresses
  4. L17:642 strongly suppresses
  5. L18:713 strongly suppresses
  6. L4:295 strongly suppresses
  7. L16:432 strongly suppresses period after month or initial
  8. L5:168 strongly suppresses
  9. L19:831 suppresses
  10. L3:595 suppresses
  11. L3:871 suppresses
  12. L3:259 suppresses
  13. L19:2497 suppresses
  14. L14:2364 suppresses
  15. L14:422 suppresses periods after initials and abbreviations
  16. L5:1257 suppresses
  17. L5:1653 suppresses
  18. L5:1304 suppresses
  19. L6:71 suppresses
  20. L5:454 excites
  21. L17:1218 excites
  22. L6:192 excites
  23. L16:844 suppresses
  24. L8:541 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 -4.49 and -4.341 values between -3.33 and -3.191 values between -2.75 and -2.611 values between -2.61 and -2.462 values between -2.17 and -2.031 values between -1.88 and -1.741 values between -1.45 and -1.302 values between -1.16 and -1.012 values between -1.01 and -0.878 values between -0.87 and -0.728 values between -0.72 and -0.5814 values between -0.58 and -0.4424 values between -0.44 and -0.2957 values between -0.29 and -0.15181 values between -0.15 and -0.002200 values between -0.00 and 0.151217 values between 0.15 and 0.29165 values between 0.29 and 0.4321 values between 0.43 and 0.581 values between 1.16 and 1.30offnegativepositiveoften

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. range
  2. inf
  3. ivant
  4. segue
  5. ronic
  6. stretch
  7. lude
  8. hf
  9. rb
  10. utches

Suppresses

  1. Gutenberg
  2. cter
  3. -binary
  4. śmy
  5. SizePolicy
  6. .toLocale
  7. 不理
  8. 这两个
  9. ForRow
  10. },{"