Neuron Explorer

Qwen3-0.6B

unclear: parentheses and word fragments

A neuron in layer 15 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron is hard to pin down. It fires on a mix of opening brackets before numbers and on the middle pieces of split words, such as the 'ated' in 'inaugurated'. Turning it up made the model produce repetitive or unrelated text.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    ay, lists just over 1,500 names on 23 bronze plaques. It was erected in 1923, and inaugurated by the then landowner, Lord Leverhulme. The first memorial in

  2. #2 peak

    1 problem (12 points). Part B has 5 shorter problems (7 points each, so 35 points). Part C has 3 traditional problems (15 points each so 45 points). Total is 92 points

  3. #3 peak

    ology for hints that the Red Planet may have ever had conditions hospitable to life.⏎The startlingly clear photo, one of several new views released by NASA today (Aug. 8), shows Curiosity's heat shield after it was shed about

  4. #4 peak

    temperature for the year in Mount Surprise is 73.0°F (22.8°C). The warmest month, on average, is November with an average temperature of 81.0°F (27.2°C).

  5. #5 peak

    Mosquito Bite⏎|Duration Of The Events⏎|Symptoms of dengue⏎|Three to four days⏎|Recovery from dengue⏎|Ten to twelve days⏎After the mosquito bite, the person feels standard for a

  6. #6 peak

    even though the practice has been illegal for almost 40 years.⏎It is estimated more than four million of them are children.⏎The BBC's Humphrey Hawksley reports from the village of Pipalguda in the Indian state of Orissa.

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 first one, the one that is a little more than a mile from the Earth's surface, the one that is at latitude, north, south, or... the Earth's, at latitude 0, 90°,

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

    in the Ussr. StudyMode.com. Retrieved 05, 2012, from http://www.studymode.com/essays/The-Struggle-For-Power-In-The-99432

  2. #2 peak

    remote control for situations that the machine can’t quite do things,” he said.⏎For the rest of this article, click here: http://www.abc.net.au/news/rural/2017-05-23/the-f

  3. #3 peak

    European health officials reported 1,073 cases of measles from 15 countries in Jan. 2018. Greece and France reported the most cases with 431

  4. #4 peak

    miles. The cumulative estimate for the year is 2,972.3 billion vehicle miles.⏎Comparing year-over-year figures, the West showed the most growth at 3.2 percent, followed by the South Atlantic at 2

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. L13:2934 strongly suppresses
  2. L12:408 suppresses
  3. L14:1435 excites
  4. L13:2915 suppresses
  5. L13:2125 suppresses
  6. L14:840 suppresses
  7. L13:2651 excites
  8. L13:2288 excites
  9. L14:258 excites
  10. L12:2797 excites
  11. L13:638 excites
  12. L14:241 suppresses
  13. L13:1740 excites
  14. L13:513 excites
  15. L12:2801 excites percentages and time spans in statistics
  16. L13:2748 suppresses
  17. L14:2939 excites
  18. L13:1710 excites
  19. L12:888 suppresses
  20. L14:1895 suppresses
  21. L12:2128 suppresses
  22. L13:1705 suppresses
  23. L10:1840 suppresses
  24. L14:1912 weakly suppresses document end, then unrelated text

Feeds later layers

  1. L16:2306 strongly suppresses
  2. L18:2739 suppresses
  3. L19:1481 suppresses
  4. L18:1501 suppresses
  5. L25:360 suppresses
  6. L16:431 excites
  7. L23:224 suppresses
  8. L16:2161 suppresses
  9. L20:758 suppresses
  10. L16:1680 suppresses
  11. L26:13 excites
  12. L26:234 suppresses
  13. L18:2014 suppresses
  14. L17:1878 suppresses
  15. L19:1947 suppresses
  16. L20:521 suppresses
  17. L19:1921 suppresses
  18. L27:218 suppresses
  19. L16:2759 suppresses
  20. L18:291 excites
  21. L19:1732 suppresses
  22. L16:508 suppresses
  23. L16:1165 weakly suppresses
  24. L20:947 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 -5.21 and -5.041 values between -2.40 and -2.231 values between -2.05 and -1.881 values between -1.88 and -1.701 values between -1.70 and -1.523 values between -1.35 and -1.172 values between -1.17 and -1.007 values between -1.00 and -0.829 values between -0.82 and -0.6522 values between -0.65 and -0.4760 values between -0.47 and -0.30270 values between -0.30 and -0.121509 values between -0.12 and 0.061171 values between 0.06 and 0.23631 values between 0.23 and 0.41163 values between 0.41 and 0.5840 values between 0.58 and 0.766 values between 0.76 and 0.933 values between 0.93 and 1.115 values between 1.11 and 1.291 values between 1.46 and 1.641 values between 1.64 and 1.81offnegativepositiveoften

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. .Reverse
  2. lest
  3. .Serial
  4. .SplitContainer
  5. VERIFY
  6. )::
  7. ental
  8. erin
  9. /design
  10. ircuit

Suppresses

  1. (
  2. (
  3. (-
  4. ($
  5. (
  6. ({
  7. ($_
  8. 授
  9. (**
  10. (.