Neuron Explorer

Qwen3-0.6B

comma after a full date

A neuron in layer 22 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 the comma in dates like 'Jan 09, 2017' or 'July 12, 2017', the spot right before the year. It goes quiet, or flips negative, on the middle digits of years written inside text, such as the '3' in 1380. Turning it up mostly made the model fall into loops of blank lines, so the steering results do not tell us much more.

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

    Jan 09, 2017 08:57 AM EST⏎MIT Professor

  2. #2 peak

    direction you would get more water pressure (voltage) because the pumps are helping each other<|endoftext|>Canine fear-related aggression (Proceedings)⏎Nov 01, 2009⏎CVC IN SAN DIEGO PROCEEDINGS⏎

  3. #3 peak

    pale little girl with a thin neck, and spindling legs half-hidden in the folds of Aunt Frances’s skirts. But she didn’t<|endoftext|>Dec 6, 2023⏎Airbus has successfully passed the Preliminary Design Review

  4. #4 peak

    York Times, Project Glass, technology, turn by turn directions, UI, User interface, video chat, Virtual reality, Virtual reality glasses⏎January 18, 2012 • 11:07 am

  5. #5 peak

    with something familiar for scale pic.twitter.com/OdQA07qhBb⏎— Adrian Luckman (@adrian_luckman) July 12, 2017⏎You may imagine that the breaking off of the iceberg

  6. #6 peak

    every bit of opinion you might start writing, think of including an “and…” statement.⏎— Danielle Degelman (@deedegs) November 26, 2014⏎One approach if giving criticism is to put it inside

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: the family of a family with a big family, and a family with a small family. If you have a large family, there is a big family with a big family, and a small family with a large family. The pattern is a

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

    , is thought to be the name or epithet of a Kartid amir or military commander. Destroyed a second time by Timur’s army (1380), Qal’a-ye Ikhtiyar al-Din was

  2. #2 peak

    the pope's favor forced him to leave Rome and work for King Alfonso of Naples (1452-1455), but he resumed his position as papal

  3. #3 peak

    emerged as the leader of local Turks in the fight against the tottering Byzantine Empire. The final conquest of the Byzantines was not achieved until 1453 with the fall of Constantinople (now Istanbul), but by that date

  4. #4 peak

    Colov.⏎Later, the tower had a defensive character as well, especially in the time of the attack of Omer Pasha on Montenegro in 1862. At the end of the 19th century this building was

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. L18:3021 excites
  2. L14:2049 weakly excites
  3. L20:151 weakly excites
  4. L17:1333 weakly excites
  5. L19:1272 weakly suppresses
  6. L19:1308 weakly suppresses
  7. L9:42 weakly suppresses
  8. L21:1042 weakly excites
  9. L7:1564 weakly excites website boilerplate and navigation text
  10. L12:1443 weakly excites opening parenthesis after a name or term
  11. L12:1145 weakly excites
  12. L10:80 weakly suppresses
  13. L2:1049 weakly excites
  14. L12:1140 weakly excites
  15. L14:3017 weakly excites
  16. L18:901 weakly excites
  17. L5:2915 weakly suppresses
  18. L12:668 weakly excites
  19. L9:2659 weakly suppresses
  20. L15:23 weakly suppresses
  21. L1:2161 weakly suppresses
  22. L4:460 weakly suppresses
  23. L5:1081 weakly suppresses
  24. L14:1811 weakly excites word before an evaluative adjective in essays

Feeds later layers

  1. L23:3020 excites
  2. L23:1093 excites
  3. L26:555 excites
  4. L27:1779 excites
  5. L23:233 suppresses
  6. L23:337 weakly excites
  7. L25:487 weakly suppresses
  8. L25:1667 weakly excites
  9. L23:910 weakly excites
  10. L24:930 weakly excites
  11. L25:465 weakly excites
  12. L23:550 weakly excites
  13. L24:274 weakly excites unclear: mixed technical tokens
  14. L24:1601 weakly suppresses
  15. L24:2744 weakly excites
  16. L25:459 weakly suppresses
  17. L27:1452 weakly excites
  18. L23:61 weakly suppresses line breaks mid-sentence
  19. L26:2078 weakly excites
  20. L27:112 weakly suppresses
  21. L27:454 weakly excites
  22. L27:825 weakly excites
  23. L27:2187 weakly excites
  24. L23:113 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 -3.98 and -3.801 values between -3.80 and -3.621 values between -3.62 and -3.451 values between -3.27 and -3.091 values between -2.73 and -2.561 values between -2.02 and -1.842 values between -1.84 and -1.661 values between -1.66 and -1.492 values between -1.49 and -1.311 values between -1.31 and -1.134 values between -1.13 and -0.958 values between -0.95 and -0.7728 values between -0.77 and -0.5955 values between -0.59 and -0.42162 values between -0.42 and -0.24635 values between -0.24 and -0.062295 values between -0.06 and 0.12446 values between 0.12 and 0.30154 values between 0.30 and 0.4867 values between 0.48 and 0.6525 values between 0.65 and 0.837 values between 0.83 and 1.016 values between 1.01 and 1.191 values between 1.19 and 1.371 values between 2.44 and 2.611 values between 2.61 and 2.791 values between 2.97 and 3.15offnegativepositiveoften

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. (⏎⏎
  3. [⏎
  4. ь
  5. (⏎
  6. ьев
  7. [⏎⏎
  8. URN
  9. 之人
  10. (⏎

Suppresses

  1. MSC
  2. Rnd
  3. ]={
  4. ={
  5. ={}⏎
  6. @register
  7. ContextHolder
  8. CEL
  9. NSE
  10. →