Neuron Explorer

Qwen3-0.6B

the digit 6 in years 19xx

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 fires on the middle digit of years in the 1900s, especially the 6 in 1961, 1963 or 1969. Turning it up did not make the text talk about years, and the link to the negative side is unclear.

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

    would be a combination of the then-Egyptian anthem and "Homat elDiyar" When Syria seceded from the union in 1961, the⏎anthem was completely restored.⏎Translation of Syrian National Anthem:⏎

  2. #2 peak

    , pondering the influence of the various strands of Irish nationalism on the modern state.⏎Finally, the book reviews the conflict in Northern Ireland from the 1960s all the way to the fall of

  3. #3 peak

    17-1918), the Second World War (1941-45), and events such as the Cuban Missile Crisis (1963).⏎During the Seminole Wars, the Spanish American War,<|endoftext|>Could

  4. #4 peak

    and passed away after 6 years.⏎Dolly wasn’t the first vertebrate to be cloned. That honour belongs to a carp, cloned in 1963.⏎The first cloned mouse,

  5. #5 peak

    in the war. This represents about 22% of the men of fighting age, a terrible loss in a small community. Returning servicemen formed a Kincardine O’Neil branch of the Royal British legion. This link will,

  6. #6 peak

    that space travel was beyond the scope of human capacity, yet astonishingly, technological advances allowed a human to walk on the surface of the moon in 1969 – a feat unimaginable to humans living a mere century before. Since that

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: through the use of through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through through

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

    . But when you consider that stress has been linked to a host of other disorders, including depression, heart attacks, hypertension and increased susceptibility to infection, it’s easy to understand how one herb can have an all-encompassing, tonic effect.⏎

  2. #2 peak

    Canada Metals (Pacific) Limited, Delta, BC, Canada (Martyr Brand), www.canmet.com<|endoftext|>5 interesting facts about the Singapore River clean-up⏎Published on Jul

  3. #3 peak

    DOES SUN EXPOSURE CAUSE CANCER?⏎One of the reasons people don’t get as much sun as they

  4. #4 peak

    . In fact, the sequence has 30 conjectured members. Similarly, our estimate predicts 2.5 members starting from 86. It is easy to check that the sequence doesn’t contain any more numbers below 200

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. L19:1473 excites "can" after a noun in a clause
  2. L4:173 suppresses
  3. L17:953 excites the word 'can' after a subject
  4. L20:2487 suppresses
  5. L3:978 suppresses
  6. L10:384 excites
  7. L18:1491 suppresses
  8. L19:2701 suppresses
  9. L19:243 weakly excites
  10. L19:1865 weakly suppresses
  11. L19:1138 weakly excites
  12. L21:443 weakly suppresses
  13. L4:904 weakly excites
  14. L16:1153 weakly suppresses
  15. L20:2375 weakly excites
  16. L4:777 weakly suppresses
  17. L17:670 weakly excites
  18. L21:381 weakly suppresses zeros inside long number codes
  19. L1:1926 weakly excites
  20. L5:2772 weakly excites
  21. L2:2386 weakly suppresses
  22. L2:1259 weakly suppresses
  23. L10:2291 weakly suppresses
  24. L16:2702 weakly suppresses

Feeds later layers

  1. L24:883 weakly suppresses
  2. L23:756 weakly suppresses
  3. L27:219 weakly excites
  4. L26:2729 weakly excites
  5. L23:833 weakly suppresses
  6. L23:2441 weakly excites
  7. L23:1687 weakly suppresses
  8. L23:2735 weakly suppresses
  9. L25:523 weakly suppresses
  10. L24:2435 weakly excites
  11. L24:2654 weakly excites
  12. L26:3027 weakly suppresses
  13. L23:1061 weakly excites
  14. L24:921 weakly suppresses
  15. L26:1915 weakly excites
  16. L23:2161 weakly suppresses
  17. L25:2479 weakly suppresses
  18. L26:1832 weakly suppresses
  19. L27:2054 weakly excites
  20. L27:686 weakly excites
  21. L24:2846 weakly excites
  22. L26:1136 weakly suppresses
  23. L24:1605 weakly suppresses
  24. L27:1802 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 -46.41 and -44.471 values between -25.09 and -23.162 values between -11.53 and -9.595 values between -9.59 and -7.665 values between -7.66 and -5.727 values between -5.72 and -3.7825 values between -3.78 and -1.842526 values between -1.84 and 0.091294 values between 0.09 and 2.0317 values between 2.03 and 3.9711 values between 3.97 and 5.916 values between 5.91 and 7.843 values between 7.84 and 9.782 values between 9.78 and 11.721 values between 13.66 and 15.591 values between 23.34 and 25.281 values between 29.16 and 31.09offnegativepositiveoften

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. zy
  2. law
  3. icles
  4. ncia
  5. Tet
  6. -collar
  7. -law
  8. Braz
  9. law
  10. nc

Suppresses

  1. 通过
  2. through
  3. -through
  4. _through
  5. 以
  6. 见
  7. 看你
  8. Through
  9. through
  10. pector