Neuron Explorer

Qwen3-0.6B

unclear: fragments of rare nouns

A neuron in layer 2 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 pieces of uncommon words, like the middle of 'puffer' fish, 'tabbouleh' and 'tuberous'. The pattern is not very clear, and turning it up gave rambling text with no obvious theme.

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

    Connecting⏎- 2 The Network⏎- 3 Printing Files⏎- 4 GuardBuddies⏎- 5 Radio⏎- 6 Troubleshooting⏎- 7 Tapping the Wire⏎- 8 Scripting⏎

  2. #2 peak

    like minerals and vitamins.⏎Examples of salads: Tabbouleh, fattoush, rocca salad, greek salad.⏎2.

  3. #3 peak

    Southeast Asia, and Africa. They are primarily found in the submerged plants along the riverside.⏎There are at least about 150 known species of puffer. Some common species are,⏎Red Eye Puffer⏎Dwarf Pea P

  4. #4 peak

    the large four teeth used for crushing the hard shells of their prey. They are slow-moving fish and show dorsal, anal, caudal fin motion.⏎Puffer fish are common in tropical and temperate climates. They mostly live in brack

  5. #5 peak

    would pass away before<|endoftext|>The Badrouzi Festival or “Kajin” was held in ancient Iran in honor of

  6. #6 peak

    corn oil and safflower oil may also self-heat. Fish oils are notorious for their self-heating properties. Mineral oils like white spirit, mineral turpentine or lubricating oil are not prone to self-heating and will not cause

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 been in the hospital for a week, and I have been told by the nurse. I have been told by the nurse, and I have been in the hospital for a week. I have been in the hospital for a week,

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

    PA. 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

  2. #2 peak

    or she is just a huge airhead (this is my view). She truly seems not to know that<|endoftext|>Three Meetings in June 2006 Underscore UN Efforts to Improve Early Warning and Mitigation of Natural Disasters⏎

  3. #3 peak

    , an independent think-tank and watchdog over the environmental consequences of Canadian government and corporate activities around the world. Her books include In the Name of Progress: The Underside of Foreign Aid, (Doubleday 19

  4. #4 peak

    has these capabilities in principle, we can go to the testing stage where we begin to collect data from real epilepsy patients.⏎“Provid<|endoftext|>Dashes vs. Underscores in URLs⏎The debate over whether to use dashes or underscores to represent spaces

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. L1:1464 excites
  2. L1:22 weakly excites
  3. L0:53 weakly suppresses
  4. L0:1224 weakly suppresses
  5. L1:3066 weakly suppresses
  6. L1:2683 weakly excites
  7. L1:1757 weakly suppresses
  8. L0:1301 weakly suppresses
  9. L1:10 weakly excites
  10. L0:182 weakly excites
  11. L0:1342 weakly suppresses
  12. L1:1029 weakly excites
  13. L0:178 weakly suppresses
  14. L1:2801 weakly excites
  15. L1:162 weakly excites
  16. L1:2342 weakly excites
  17. L1:1583 weakly excites
  18. L1:1313 weakly excites
  19. L1:619 weakly suppresses
  20. L0:404 weakly excites
  21. L0:102 weakly suppresses
  22. L1:491 weakly excites
  23. L1:1250 weakly excites
  24. L0:1911 weakly suppresses

Feeds later layers

  1. L4:2383 suppresses
  2. L4:2820 suppresses
  3. L18:805 suppresses
  4. L19:1206 suppresses
  5. L16:1558 suppresses
  6. L20:2017 weakly suppresses
  7. L17:2510 weakly excites
  8. L3:232 weakly suppresses 'on' after a noun like 'impact on'
  9. L21:1246 weakly suppresses
  10. L17:1369 weakly suppresses "of" after is/are/be (of + quality)
  11. L8:1945 weakly excites
  12. L14:233 weakly suppresses
  13. L11:2865 weakly excites
  14. L5:977 weakly excites
  15. L17:1196 weakly suppresses
  16. L8:1960 weakly suppresses
  17. L4:925 weakly excites
  18. L12:1378 weakly suppresses
  19. L17:2438 weakly suppresses
  20. L15:936 weakly excites
  21. L17:2615 weakly suppresses
  22. L7:2133 weakly suppresses
  23. L9:246 weakly excites
  24. L22:2679 weakly suppresses cycles and periods, before the word 'cycle'

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 -3.37 and -3.251 values between -2.07 and -1.952 values between -1.71 and -1.601 values between -1.60 and -1.481 values between -1.24 and -1.121 values between -1.00 and -0.882 values between -0.88 and -0.771 values between -0.77 and -0.652 values between -0.65 and -0.533 values between -0.53 and -0.417 values between -0.41 and -0.2992 values between -0.29 and -0.17745 values between -0.17 and -0.062219 values between -0.06 and 0.06723 values between 0.06 and 0.1893 values between 0.18 and 0.3010 values between 0.30 and 0.421 values between 0.42 and 0.542 values between 0.54 and 0.661 values between 1.25 and 1.37offnegativepositiveoften

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. Departments
  2. offices
  3. retweeted
  4. 攘
  5. modules
  6. Laundry
  7. institutions
  8. gy
  9. tongues
  10. Replay

Suppresses

  1. pace
  2. åde
  3. chron
  4. =sub
  5. ówi
  6. intermittent
  7. }}⏎⏎
  8. �
  9. siege
  10. iseconds