Neuron Explorer

Qwen3-0.6B

unclear: table pipes and split word fragments

A neuron in layer 19 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

It mostly lights up on the vertical bar that starts a new row in a text table, and on pieces of split words or web addresses. Turning it up or down did not give a clear 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

    i and add -es.⏎If a noun ends in vowel + y, add -s. Words ending in -quy don’t follow this rule (as in soliloquies).⏎Add -s to most nouns ending in -f.⏎

  2. #2 peak

    ⏎|Pacific Coast Tick⏎||Light brown with a spotted white collar behind head⏎||Pacific Ocean coastal areas⏎||Colorado Tick Fever, Tick Paralysis⏎|Rocky Mountain Wood Tick⏎||Tear-shaped, black body with brown and

  3. #3 peak

    ut-shaped and covered with small spines⏎||Causes intense irritation, rubbing, and hair loss⏎|Tropical Horse Tick⏎||Dark black body⏎|Western Black-Legged<|endoftext|>Your garden is an orchestra! Healthy gardens play to

  4. #4 peak

    ar flow cabinets, and other biological safety cabinet components protect.⏎Read More @ https://cmibloginsight.blogspot.com/2023/05/biological-safety-cabinets-ensuring.html<|endoftext|>World

  5. #5 peak

    done.⏎A practical guide to avoiding dietary disasters⏎Tortoise Trust guidelines for:⏎Feeding Leopard tortoises - Feeding Mediterranean tortoises - Feeding Redfoot and Yellowfoot tortoises - Feeding Wood turtles and Box turtles

  6. #6 peak

    and Dashkat Hakaayuʔ “Whitebread” (who told 19% of the stories) https://archive.org/details/traditionsofcadd00dorsrich⏎Explore more about the Caddo history and

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 same amount of information can be represented as a math (recall that all math can represent any of the represented sequences of numbers, and the same math (recall that all math can represent any of the representable sequences of numbers)

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

    borne by Al-H<|endoftext|>If you were asked to name the countries of the world that are most strongly identified with the Roman Church, Poland would surely be one that came to mind. Pope John Paul II is a notable representative of Polish faith. Yet

  2. #2 peak

    around here, I’ve not tried to grow them in the garden. But Sara Stein had high praise for them as garden plants in her book Noah’s Garden:⏎Blueberries are happy

  3. #3 peak

    from the good old Spring and Autumn period (772BC to 403BC) and the Warring States period right after (403BC to

  4. #4 peak

    is inexorably linked with silver, this metal that provoked a demographic and economic boom in the city during three centuries, as well as its decline and poverty that came next when it became rarer. The richness of the region’s basements gave

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. L15:958 excites
  2. L14:2726 suppresses
  3. L18:573 excites
  4. L18:452 suppresses
  5. L16:1315 excites word or phrase being named or defined
  6. L15:557 excites
  7. L14:1396 suppresses word before a verb of action or effect
  8. L18:706 excites
  9. L17:265 suppresses
  10. L18:569 suppresses
  11. L18:1165 suppresses
  12. L13:527 suppresses
  13. L14:2918 excites
  14. L14:313 excites
  15. L16:2365 suppresses
  16. L18:489 suppresses
  17. L17:2749 excites
  18. L14:2651 weakly suppresses
  19. L18:97 weakly excites
  20. L14:826 weakly excites
  21. L16:786 weakly excites
  22. L13:330 weakly excites
  23. L15:1310 weakly excites
  24. L18:928 weakly suppresses

Feeds later layers

  1. L21:1227 excites
  2. L20:697 excites
  3. L24:174 weakly excites
  4. L21:2397 weakly excites
  5. L20:795 weakly suppresses
  6. L21:1467 weakly suppresses
  7. L20:2129 weakly excites
  8. L23:2778 weakly suppresses
  9. L20:1963 weakly excites
  10. L21:2059 weakly suppresses
  11. L20:958 weakly excites
  12. L22:1451 weakly suppresses
  13. L25:215 weakly excites
  14. L21:831 weakly excites
  15. L22:2493 weakly suppresses
  16. L21:2055 weakly suppresses
  17. L22:1140 weakly suppresses
  18. L21:2231 weakly excites
  19. L25:1176 weakly excites
  20. L21:58 weakly excites
  21. L23:159 weakly excites
  22. L24:1002 weakly suppresses
  23. L26:1997 weakly suppresses
  24. L23:47 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 -8.48 and -8.031 values between -7.58 and -7.132 values between -7.13 and -6.681 values between -6.68 and -6.235 values between -6.23 and -5.785 values between -5.78 and -5.332 values between -5.33 and -4.887 values between -4.88 and -4.4319 values between -4.43 and -3.9824 values between -3.98 and -3.5344 values between -3.53 and -3.0873 values between -3.08 and -2.6397 values between -2.63 and -2.18155 values between -2.18 and -1.73223 values between -1.73 and -1.28374 values between -1.28 and -0.83601 values between -0.83 and -0.381023 values between -0.38 and 0.07579 values between 0.07 and 0.52257 values between 0.52 and 0.97144 values between 0.97 and 1.4297 values between 1.42 and 1.8760 values between 1.87 and 2.3246 values between 2.32 and 2.7716 values between 2.77 and 3.2216 values between 3.22 and 3.6712 values between 3.67 and 4.1210 values between 4.12 and 4.575 values between 4.57 and 5.021 values between 5.02 and 5.473 values between 5.47 and 5.921 values between 5.92 and 6.371 values between 6.37 and 6.821 values between 6.82 and 7.271 values between 7.72 and 8.171 values between 9.07 and 9.52offnegativepositiveoften

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. 样
  2. 벌
  3. 个城市
  4. passports
  5. ercise
  6. entities
  7. athletes
  8. Celtics
  9. imately
  10. verity

Suppresses

  1. icular
  2. isen
  3. tern
  4. 靛
  5. ]=[
  6. plat
  7. cob
  8. etration
  9. isFunction
  10. prech