Neuron Explorer

Qwen3-0.6B

unclear: function words in descriptive prose

A neuron in layer 21 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 gives weak responses to small connecting words like 'which', 'a', 'with' and 'becomes' in science or health explanations. Its negative side fires inside technical terms such as 'pesticides' or 'Cyber Security'. Turning it down made the model repeat 'in' endlessly, but the pattern itself is not clear.

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

    stick refers to an incomplete fracture. Unlike other types of broken bones, a greenstick fracture isn’t fully separated.⏎OpenAn open fracture is a bone break in which the bone penetrates and protrudes through the skin.⏎Treating Broken BonesTre

  2. #2 peak

    bone.⏎ObliqueAn oblique fracture runs diagonally across the bone⏎ComminutedA comminuted bone break is present in at least three parts with multiple fractures.⏎SegmentalA segmental fracture refers to one bone with fractures in two

  3. #3 peak

    in fresh water.⏎A wide variety of sponges exist, ranging in size from 1 inch to 4 feet in diameter. Inside a sponge there is a system of openings and passages called a canal system. The sponge is nourished by

  4. #4 peak

    older people. This disease affects the tissue covering the ends of bones in a joint (cartilage).In a person with osteoarthritis, the cartilage becomes damaged and worn out causing pain, swelling, stiffness and restricted movement

  5. #5 peak

    assesses word meaning within the context of particular passages. Students are asked to demonstrate their understanding of words by recognizing what meaning the word contributes to the passage in which it appears.”⏎For more information, The report card can be found online: www.n

  6. #6 peak

    the nail. The mottled light and dark tissue (2) is the top part of the nail.⏎As you cut the nail deeper, you will see a homogeneous gray to pink oval (3) starting to appear at the top of the

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: in the in the in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in

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

    in Produce,⏎Baby Food, Tap Water⏎Washington, D.C. – Environmental Working Group has released the eighth edition of its Shopper’s Guide to Pesticides in Produce with updated information on 45 popular fruits and vegetables and their total

  2. #2 peak

    few recombine, the energy required in maintaining the plasma will<|endoftext|>Welcome to the Teaching Tolerance blog, a place where educators who care about diversity, equity and justice can find news, suggestions, conversation and support.⏎The big day is just around

  3. #3 peak

    due to these security breaches is over 10.1 million dollars.⏎Thus cybersecurity remains a top concern and challenge for the healthcare industry.⏎Impacts of Cyber Security Threats⏎The results of a cyber-attack on a provider network have four

  4. #4 peak

    this article, click here: http://www.abc.net.au/news/rural/2017-05-23/the-future-of-automation-in-the-mining-industry/8550636<|endoftext|>Service

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. L13:1778 suppresses "of" or "on" after part, face or footprint
  2. L18:2598 suppresses
  3. L20:1432 excites
  4. L15:1744 suppresses
  5. L11:2651 weakly suppresses
  6. L16:2558 weakly excites
  7. L12:2938 weakly excites
  8. L14:2229 weakly suppresses
  9. L1:2199 weakly suppresses
  10. L17:1256 weakly suppresses
  11. L9:2040 weakly suppresses
  12. L10:3010 weakly excites
  13. L9:2929 weakly excites
  14. L9:2531 weakly excites
  15. L15:2987 weakly excites
  16. L16:2391 weakly suppresses
  17. L4:1131 weakly excites
  18. L14:1770 weakly excites
  19. L18:1654 weakly suppresses
  20. L12:2077 weakly excites
  21. L10:2649 weakly excites
  22. L19:241 weakly excites
  23. L17:1319 weakly suppresses
  24. L9:2700 weakly excites

Feeds later layers

  1. L24:893 strongly suppresses
  2. L26:207 strongly excites
  3. L26:825 strongly suppresses
  4. L27:405 strongly suppresses
  5. L23:370 strongly suppresses
  6. L24:116 excites
  7. L26:584 suppresses
  8. L23:476 suppresses
  9. L22:439 excites
  10. L26:786 excites
  11. L24:1284 suppresses
  12. L25:1790 suppresses the digit 1 starting a statistic
  13. L26:1478 excites
  14. L24:887 excites
  15. L26:164 excites
  16. L26:1876 weakly suppresses
  17. L25:2282 weakly suppresses
  18. L23:1366 weakly excites
  19. L22:2430 weakly excites
  20. L23:1784 weakly suppresses
  21. L23:2514 weakly excites
  22. L22:1110 weakly excites
  23. L22:1747 weakly excites
  24. L22:2591 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 -7.52 and -7.142 values between -7.14 and -6.762 values between -6.38 and -6.004 values between -6.00 and -5.622 values between -5.62 and -5.241 values between -5.24 and -4.863 values between -4.48 and -4.103 values between -4.10 and -3.724 values between -3.72 and -3.344 values between -3.34 and -2.966 values between -2.96 and -2.589 values between -2.58 and -2.2013 values between -2.20 and -1.8218 values between -1.82 and -1.4534 values between -1.45 and -1.0772 values between -1.07 and -0.69409 values between -0.69 and -0.311600 values between -0.31 and 0.07785 values between 0.07 and 0.45374 values between 0.45 and 0.83211 values between 0.83 and 1.21111 values between 1.21 and 1.5981 values between 1.59 and 1.9749 values between 1.97 and 2.3550 values between 2.35 and 2.7316 values between 2.73 and 3.1121 values between 3.11 and 3.499 values between 3.49 and 3.872 values between 3.87 and 4.253 values between 4.25 and 4.633 values between 4.63 and 5.012 values between 5.01 and 5.392 values between 5.39 and 5.771 values between 6.53 and 6.911 values between 7.29 and 7.67offnegativepositiveoften

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. erty
  2. imentary
  3. IMO
  4. roy
  5. elijk
  6. iero
  7. ieri
  8. тер
  9. ße
  10. oria

Suppresses

  1. 在
  2. in
  3. 在他的
  4. 在一个
  5. 在一
  6. 在学校
  7. 在他
  8. 在网络
  9. 在香港
  10. 在日本