Neuron Explorer

Qwen3-0.6B

the 9 in years like 19xx

A neuron in layer 20 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 fires on the middle digit of years from the 1900s, such as the 9 in 1998 or 1993, when the number appears inside a citation or a date. Its most negative moments are on the start of the word 'statue', which is a different pattern. Turning it up or down changed the text in ways that did not match the pattern, so the steering results tell us little.

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

    who are expected to convey these values to learners formally through the implementation of values education. The concept of values education will be explained below.⏎Robb (1998) defines values education as an activity whereby people are educated about aspects that determine

  2. #2 peak

    out landmark studies on the natural history of cervical cancer that firmly established HPV as the necessary cause of this malignancy (Schiffman et al., 1993, 2007). Their efforts laid the groundwork for vaccine development

  3. #3 peak

    National Medal of Science (1986); the APA’s Award for Outstanding Lifetime Contributions to Psychology (1993)⏎- ACM fellow (1994)⏎- IJCAI Award for Research Excellence (1995

  4. #4 peak

    ivian Army and the CIA in 1967 (Castaneda 1998, 326).⏎What Individuals Thought of Che Guevara⏎

  5. #5 peak

    persuasiveness of viral campaigns in social network sites: Persuasive intent as the underlying mechanism. Journal of Marketing, 18(1), 39-53. http://dx.doi.org/10.10

  6. #6 peak

    the Society for American Baseball Research.⏎Jerry Malloy (1946-2000) was a pioneer researcher who was honored in 1997 by the creation of an annual Negro League Conference named for him. He was

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 a 48-cm (about 18-inch) machine with a 300-meter wave machine that is in a state of "re-connection." I have had a lot of learning and growth from this

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

    Manila. The bridge was rebuilt in spartan steel and concrete. None of the statuary of mermen on dolphin and allegorical sculptures are present in today’s Jones

  2. #2 peak

    . In June 47 B.C., Cleopatra bore a son, whom she claimed was Caesar's and named Caesarion, meaning "little<|endoftext|>The statuette is representing a Nubian kneeling pharaoh making an offering to the

  3. #3 peak

    of essential liturgical equipment of a sanctuary of Egypt or Nubia.⏎Solid copper alloy statuette formely clad with in precious-metal leaf.⏎On his pleated k

  4. #4 peak

    the entire state of Delaware. The iceberg goes to the depth of 600 and 700 feet below the sea which would mean stacking up two Statues of Liberty one over another.⏎The humungous piece of ice floating in

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. L18:1628 strongly excites
  2. L17:166 strongly excites
  3. L16:43 excites
  4. L18:541 excites
  5. L1:177 suppresses
  6. L2:1611 suppresses
  7. L0:1816 excites
  8. L18:1104 suppresses
  9. L5:204 excites
  10. L19:183 excites
  11. L10:233 suppresses
  12. L18:440 suppresses
  13. L7:287 excites
  14. L18:2397 excites 'followed' and 'searched' after a subject
  15. L2:2497 excites
  16. L16:673 excites
  17. L10:951 suppresses
  18. L1:5 suppresses
  19. L16:76 excites
  20. L18:569 excites
  21. L10:722 suppresses
  22. L19:381 suppresses word-start 'acc' fragments
  23. L16:30 suppresses
  24. L2:1585 excites

Feeds later layers

  1. L21:334 strongly suppresses
  2. L26:674 suppresses
  3. L23:2702 suppresses
  4. L24:2482 suppresses
  5. L23:2397 suppresses
  6. L22:1598 excites
  7. L27:2265 weakly suppresses
  8. L27:2553 weakly excites capital T starting a word
  9. L23:2608 weakly suppresses
  10. L24:2452 weakly suppresses
  11. L23:2940 weakly suppresses
  12. L21:1793 weakly excites
  13. L27:2187 weakly suppresses
  14. L25:2249 weakly suppresses
  15. L25:1669 weakly suppresses
  16. L21:2731 weakly excites
  17. L25:2234 weakly suppresses
  18. L26:2435 weakly excites
  19. L27:1863 weakly excites
  20. L23:2169 weakly excites
  21. L22:975 weakly excites
  22. L21:2887 weakly suppresses
  23. L27:824 weakly excites
  24. L21:271 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 -91.44 and -89.081 values between -67.86 and -65.501 values between -56.07 and -53.711 values between -51.35 and -48.992 values between -46.63 and -44.281 values between -44.28 and -41.921 values between -37.20 and -34.841 values between -27.77 and -25.411 values between -25.41 and -23.053 values between -23.05 and -20.691 values between -20.69 and -18.341 values between -18.34 and -15.983 values between -15.98 and -13.624 values between -13.62 and -11.264 values between -11.26 and -8.905 values between -8.90 and -6.558 values between -6.55 and -4.1915 values between -4.19 and -1.832331 values between -1.83 and 0.531523 values between 0.53 and 2.89offnegativepositiveoften

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. station
  2. Goes
  3. go
  4. 两点
  5. ual
  6. (lower
  7. 脂肪
  8. ACHI
  9. /title
  10. pecial

Suppresses

  1. quo
  2. ards
  3. hread
  4. Trinity
  5. ycz
  6. onet
  7. essenger
  8. шие
  9. (optimizer
  10. roid