Neuron Explorer

Qwen3-0.6B

decade digit in years like 1930s

A neuron in layer 4 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 last digit of a decade written as a year, such as the 0 in "1930s" or "mid-1960s." It stays quiet on most other numbers. Turning it up did not change the writing in a way that matches this pattern.

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 encryption key for every transaction hasn't been an acceptable practice since the Caesar Cipher was popular in ancient Rome. Furthermore, since the mid-1970s, the DES algorithm was known to be vulnerable to brute-force attacks because of

  2. #2 peak

    The US Congress has had oversight responsibilities over the CIA since the Agency was established in 1947. However, prior to the mid-1970s, oversight responsibilities resided in the Armed Services Committees of both chambers and were

  3. #3 peak

    instruction. In the mid-1930s the central government tried to culturally assimilate the local population through the process of

  4. #4 peak

    00's, and settled in Wrightstown, Bucks County where Lacey lived with his sister.⏎Early in the 1870's, Lacey sat talking in the farm house of his friend William Kirk,

  5. #5 peak

    .C.C. and the W.P.A.—most profoundly influenced their lives.⏎Assess the impact of mass media on the American culture in the 1930s.⏎What factors contributed to the nationalization of culture? What kinds of movies

  6. #6 peak

    . military to experiment with containers. First used by military traffic commands, containers were extended to commercial use in the mid-1960s. Container use has continued to grow steadily in the United States and is projected

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 three different people who are in a group. The people are all in the group. The group is in the groups. They are in the groups. I have not added anything. I have not added anything. The group is in the

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

    windy. It is definitely worth the trouble.<|endoftext|>Most of the discussion of the accident revolves around specific problems. There was the lack of sufficient lifeboats (enough for at most 1200 on a ship carrying 220

  2. #2 peak

    of distressing actions.⏎A single act may be considered sexual harassment if the act is severe enough. The more severe the act, the fewer times it must happen for it to

  3. #3 peak

    land if dealings with a landowner have fallen short of a contract.⏎- Adverse possession: A claim of a right in land if it has been inhabited long enough. If a person has been in possession of land for 12 years,

  4. #4 peak

    scent released from a candle when it is burning.⏎Ahhhhhhhh the glorious melt pool! A full melt pool occurs when the wax has been heated long enough to reach the edge

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. L2:422 strongly suppresses
  2. L0:132 strongly excites
  3. L0:579 strongly excites
  4. L3:508 strongly excites the final zero of decades like '1980s'
  5. L2:404 suppresses
  6. L3:395 excites
  7. L2:2051 excites
  8. L1:217 excites
  9. L2:1091 excites
  10. L2:801 excites
  11. L0:2215 excites
  12. L2:1753 excites
  13. L1:2756 weakly suppresses
  14. L1:2607 weakly excites
  15. L0:82 weakly excites
  16. L3:1469 weakly suppresses
  17. L3:1763 weakly suppresses
  18. L0:2988 weakly suppresses
  19. L0:425 weakly suppresses
  20. L1:820 weakly suppresses
  21. L1:633 weakly suppresses 'that' after 'fact' or 'considering'
  22. L0:120 weakly excites
  23. L1:26 weakly excites
  24. L2:1083 weakly suppresses

Feeds later layers

  1. L8:3004 strongly excites
  2. L23:393 strongly excites
  3. L5:97 strongly excites
  4. L20:961 strongly excites
  5. L26:1008 strongly excites
  6. L27:2408 strongly excites
  7. L25:577 strongly suppresses
  8. L24:318 excites
  9. L21:690 suppresses
  10. L25:1757 suppresses
  11. L20:344 suppresses hyphen inside compound words
  12. L18:1997 suppresses
  13. L21:1145 suppresses
  14. L16:1231 excites
  15. L20:1513 suppresses
  16. L22:265 suppresses
  17. L25:231 excites
  18. L22:2060 suppresses
  19. L26:2003 suppresses hyphen joining two word parts
  20. L13:2438 excites
  21. L17:2748 excites
  22. L21:868 excites
  23. L27:2657 suppresses
  24. L25:2883 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 -0.91 and -0.834 values between -0.68 and -0.6113 values between -0.61 and -0.5329 values between -0.53 and -0.4581 values between -0.45 and -0.38174 values between -0.38 and -0.30260 values between -0.30 and -0.22400 values between -0.22 and -0.15623 values between -0.15 and -0.071356 values between -0.07 and 0.01610 values between 0.01 and 0.08170 values between 0.08 and 0.1676 values between 0.16 and 0.2429 values between 0.24 and 0.3118 values between 0.31 and 0.3915 values between 0.39 and 0.4712 values between 0.47 and 0.5410 values between 0.54 and 0.624 values between 0.62 and 0.707 values between 0.70 and 0.773 values between 0.77 and 0.854 values between 0.85 and 0.931 values between 1.00 and 1.081 values between 1.08 and 1.151 values between 1.15 and 1.232 values between 1.23 and 1.311 values between 1.38 and 1.462 values between 1.46 and 1.541 values between 2.07 and 2.15offnegativepositiveoften

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. s
  2. 们
  3. es
  4. sheets
  5. sand
  6. sam
  7. sampling
  8. sie
  9. sut
  10. ات

Suppresses

  1. ccione
  2. унк
  3. ductory
  4. getCode
  5. sgiving
  6. ViewById
  7. $("#"
  8. cheme
  9. olution
  10. ONENT