Neuron Explorer

Qwen3-0.6B

first piece of a split word

A neuron in layer 27 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 fires on the opening piece of a word that is split into pieces, such as 'spr' in sprouted or 're' in reiten. When it was turned up, the model drifted toward words like 'blowing' and 'Breaking' and their headlines.

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

    few years later<|endoftext|>Hand-coloured lithograph entitled; 'Dampfmaschinen Pferd worauf man in einer stunde von Paris nach Petersburg reiten kann' ('Steam engine horse on which one can

  2. #2 peak

    warming have been and are still being debated. A few scientists for several decades have believed that weather patterns have been affected by the electrically charged particles that the Sun sprays out as solar winds. After decades of working on that theory, a

  3. #3 peak

    4 been since making its historic space landing? Well, China’s happy to report that the seeds that were transported to the moon have sprouted! While plants have grown on the International Space Station before, this is the

  4. #4 peak

    goat, born in 1998, was actually 3 goats, all identical. Their DNA was modified so that their milk produced recombinant human antithrombin (rhAT), a protein that prevents blood from clotting.⏎

  5. #5 peak

    This ion channel is absolutely essential for releasing the ATP. If you don't have that channel, you can't taste sweet, bitter or umami."⏎The finding bolsters

  6. #6 peak

    models), and the 1965 Chevrolet Corvette Stingray (C2).⏎These brakes offer better stopping performance than comparable drum brakes, including resistance to "brake fade" caused by the overheating of brake components, and are

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: how to avoid l并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并并

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

    a highly valued veneer for its unique colors and swirly grain patterns. Far and away, most commercial Olive Ash occurs as European Ash (Fraxinus excelsior) and is typically imported from Europe.⏎Scans/Pictures: A

  2. #2 peak

    three others: Dogfeiling, Coelion and Llannerch. To relate these to the present parishes or villages, Dogfeiling spanned Aberwheeler, Llandyrnog, Llangwyfan,

  3. #3 peak

    on the East Coast of the United States, and the fifth-most-populous city in the United States. It is located in the Northeastern United States at the confluence of the Delaware and Schuylkill rivers, and it is the only

  4. #4 peak

    Glass bottles, however, also leach chemicals, including aluminum, explains a study by Clemens Reimann of the Norges Geologiske Undersøkelse in Norway and colleagues. Aluminum is sometimes added to soda-lime glass to

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. L21:1934 suppresses
  2. L26:1564 excites
  3. L24:105 suppresses
  4. L8:14 excites
  5. L0:1301 suppresses
  6. L21:237 weakly suppresses
  7. L20:2613 weakly suppresses
  8. L9:449 weakly excites
  9. L23:206 weakly suppresses
  10. L0:1224 weakly suppresses
  11. L8:7 weakly excites
  12. L8:4 weakly excites
  13. L0:178 weakly suppresses
  14. L11:2332 weakly excites
  15. L23:414 weakly suppresses
  16. L7:110 weakly suppresses
  17. L20:2406 weakly suppresses
  18. L18:550 weakly suppresses
  19. L18:569 weakly excites
  20. L8:1417 weakly excites
  21. L22:530 weakly suppresses
  22. L20:1414 weakly suppresses
  23. L7:102 weakly suppresses
  24. L8:16 weakly suppresses

Feeds later layers

None: this is the last layer.

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 -10.87 and -10.141 values between -7.23 and -6.501 values between -6.50 and -5.781 values between -5.05 and -4.321 values between -3.60 and -2.871 values between -2.87 and -2.144 values between -2.14 and -1.4158 values between -1.41 and -0.693615 values between -0.69 and 0.04202 values between 0.04 and 0.778 values between 0.77 and 1.498 values between 1.49 and 2.224 values between 2.95 and 3.681 values between 5.86 and 6.581 values between 16.04 and 16.761 values between 17.49 and 18.22offnegativepositiveoften

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. 회
  2. blown
  3. BoxLayout
  4. ра
  5. nurture
  6. 告
  7. 器官
  8. Breaking
  9. 酌
  10. .cert

Suppresses

  1. 并
  2. 并与
  3. 並
  4. lus
  5. 并通过
  6. 并于
  7. 并在
  8. iment
  9. <thead
  10. ibilidade