Neuron Explorer

Qwen3-0.6B

predicting the word two

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 seems tied to the idea of the number two. It shows up on words that lead into counting or choosing, and when it was turned up the model kept saying the word two over and over.

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

    shapes? Can you describe the⏎shapes in the picture?⏎10 space travellers are waiting to board their spaceships. There⏎are two rows of seats in the waiting room. Using the rules, where⏎are they all sitting? Can you

  2. #2 peak

    change places and the Engine is back on the main line?⏎In how many ways can you fit two of these yellow triangles⏎together? Can you predict the number of ways two blue triangles can⏎be fitted together?⏎Using different numbers of sticks

  3. #3 peak

    coastal dishes, oyster sauce, fish sauce and Clamato juice⏎Seafood, Molluscs, Crustaceans

  4. #4 peak

    , and crustaceans, such as lobsters and crabs. They are commonly found in Japanese cuisine, coastal dishes, oyster sauce, fish sauce and Clamato juice.⏎Shellfish are an excellent source of protein and vitamins and minerals

  5. #5 peak

    view it first. In case of further problems read the IDEAS help page. Note that these files are not on the IDEAS site. Please be patient as the files may be large.⏎As the access to this document is restricted, you may

  6. #6 peak

    our "Structures & Materials" activity, where children and parents will get to read the book “The True Story of the Three Little Pigs” by Jon Scieszka, build and test of the strength of structures built out of three different

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: Three great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great great

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

    fall under one of two forms of self-starvation: restricting, in which the person participates in extreme diets, fasting, or exercise; or binge-eating to purging, in which the person eats a large amount of food within a certain time

  2. #2 peak

    overwhelming research that has found multiple issues with AOUM education: censoring vital health care information, jeopardizing adolescent sexual health, stigmatizing the LGBTQ community, purporting harmful gender stereotypes and one religious perspective, and withholding information teens need to

  3. #3 peak

    on the front cover, which features a couple in classical dress in a Mediterranean countryside. The first performance of Acis and Galetea took place in the gardens at Cannons, the great house built by Handel's<|endoftext|>When school starts, sleep

  4. #4 peak

    of Cell Cycle and Growth by Extracellular Cues, Fall 2012. (Massachusetts Institute of Technology: MIT OpenCourseWare), http://ocw.mit.edu (Accessed). License: Creative Commons BY-NC-SA<|endoftext|>

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. L18:151 suppresses
  2. L19:929 excites
  3. L20:1473 suppresses
  4. L18:576 weakly suppresses
  5. L24:105 weakly suppresses
  6. L24:188 weakly excites
  7. L23:496 weakly excites
  8. L17:892 weakly suppresses
  9. L21:12 weakly excites
  10. L2:1166 weakly excites
  11. L19:1981 weakly suppresses
  12. L4:1872 weakly excites
  13. L18:819 weakly excites
  14. L16:2477 weakly excites
  15. L15:2597 weakly excites end of a web document
  16. L18:2935 weakly excites
  17. L15:1831 weakly excites
  18. L20:827 weakly suppresses
  19. L1:3030 weakly suppresses
  20. L1:2016 weakly excites
  21. L19:251 weakly excites
  22. L24:197 weakly excites
  23. L19:1877 weakly excites
  24. L13:82 weakly excites

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 -19.92 and -18.981 values between -11.41 and -10.471 values between -8.57 and -7.632 values between -7.63 and -6.681 values between -6.68 and -5.744 values between -5.74 and -4.793 values between -4.79 and -3.851 values between -3.85 and -2.905 values between -2.90 and -1.9518 values between -1.95 and -1.011130 values between -1.01 and -0.062697 values between -0.06 and 0.8829 values between 0.88 and 1.834 values between 1.83 and 2.782 values between 2.78 and 3.722 values between 3.72 and 4.671 values between 4.67 and 5.611 values between 5.61 and 6.561 values between 6.56 and 7.501 values between 10.34 and 11.291 values between 16.02 and 16.962 values between 16.96 and 17.91offnegativepositiveoften

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. 二是
  2. 二
  3. два
  4. 两种
  5. two
  6. two
  7. 第二种
  8. second
  9. ⼆
  10. 两家

Suppresses

  1. ッション
  2. thic
  3. indre
  4. 彘
  5. igraphy
  6. thr
  7. TH
  8. 三千
  9. ść
  10. fuss