Neuron Explorer

Qwen3-0.6B

'by' before a percentage or amount change

A neuron in layer 21 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the word 'by' when it introduces how much something rose, fell or changed, such as 'fallen by 79 percent' or 'cut by a quarter'. Turning it down made the model repeat 'from', which is the opposite word for change, and turning it up made the model produce percentage word problems.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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

    species assessed for population and distribution trends, 60 per cent have declined over the past 50 years; further research shows turtle dove numbers have plummeted by 80 per cent since 1995, hedgehog numbers are

  2. #2 peak

    well as at the fueling station, which also is being built and operated in partnership with NW Natural.⏎Trucks running on natural gas cut smog-producing pollutants by up

  3. #3 peak

    , blindness, kidney failure and lower-extremity amputation. Blood pressure control reduces the risk of cardiovascular disease (heart disease and stroke) among people with diabetes by between 33 percent and 50

  4. #4 peak

    WHO⏎The number of deaths from measles has fallen by 79 percent worldwide since 2000, thanks mainly to mass

  5. #5 peak

    applications. With the new composite, Suhr was able to grow the nanotubes several millimeters long. The reinforced carbon nanotube composite’s strength increased by 3,400 percent and 2,100 percent in

  6. #6 peak

    developing breast cancer and with that understanding were able to offer their conclusion.⏎"Our results suggest that losing just 5 percent or more of one's weight could cut by a quarter to a half the risk for the most common, estrogen-sensitive breast cancers,"

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: from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from from

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

    care in order to enhance their educational experience. The consequences of this lifestyle imbalance can be far-reaching. Sleep deprivation can endanger students’ physical health. Furthermore, stress resulting from being overworked and under-rested, combined with limited social interactions, can

  2. #2 peak

    closed to conserve heat, there was no exhaust ventilation in the facility, and very high levels of carbon monoxide were measured at the site.⏎Every year, workers die from carbon monoxide poisoning, usually while using fuel-burning equipment and tools in

  3. #3 peak

    aver: The working smoke detector⏎- Did you know that three out of five people who die in a fire at home have non-working or missing smoke detectors?⏎- Why do

  4. #4 peak

    Lung Association, “Tobacco kills nearly half a million Americans annually and costs the nation $333 billion per year in health-care expenses and lost productivity to boot”(Tuttle). This demonstrates that smoking causes damage to a person’s

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. L19:2091 excites
  2. L19:2707 excites
  3. L19:1335 suppresses
  4. L20:1818 suppresses
  5. L16:2011 weakly suppresses
  6. L14:3019 weakly excites
  7. L17:577 weakly excites
  8. L14:57 weakly excites
  9. L19:1644 weakly suppresses
  10. L8:1544 weakly excites
  11. L18:2019 weakly suppresses
  12. L17:729 weakly suppresses
  13. L7:831 weakly suppresses
  14. L5:1948 weakly suppresses
  15. L5:2348 weakly excites
  16. L10:2506 weakly excites
  17. L9:2245 weakly excites
  18. L17:539 weakly suppresses
  19. L11:1986 weakly suppresses
  20. L8:1133 weakly suppresses
  21. L15:1510 weakly excites
  22. L5:2533 weakly excites
  23. L9:2721 weakly excites
  24. L8:792 weakly excites

Feeds later layers

  1. L23:439 strongly suppresses
  2. L27:596 strongly suppresses
  3. L25:1206 strongly suppresses
  4. L23:1949 suppresses
  5. L22:1159 suppresses
  6. L23:2093 excites
  7. L22:2816 suppresses
  8. L22:3014 excites
  9. L22:1480 suppresses
  10. L23:1931 suppresses
  11. L24:1800 weakly excites
  12. L22:808 weakly suppresses
  13. L23:1041 weakly suppresses
  14. L25:1709 weakly suppresses
  15. L23:2142 weakly excites
  16. L23:305 weakly suppresses
  17. L23:2791 weakly excites
  18. L23:2706 weakly excites
  19. L22:1931 weakly excites
  20. L23:637 weakly suppresses
  21. L22:583 weakly suppresses
  22. L22:402 weakly excites
  23. L23:930 weakly excites
  24. L23:804 weakly suppresses

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 -26.09 and -24.981 values between -20.51 and -19.401 values between -17.17 and -16.052 values between -14.93 and -13.821 values between -13.82 and -12.701 values between -12.70 and -11.593 values between -11.59 and -10.473 values between -9.35 and -8.243 values between -8.24 and -7.124 values between -7.12 and -6.016 values between -6.01 and -4.8915 values between -4.89 and -3.7723 values between -3.77 and -2.6645 values between -2.66 and -1.54289 values between -1.54 and -0.433359 values between -0.43 and 0.69122 values between 0.69 and 1.8119 values between 1.81 and 2.924 values between 2.92 and 4.043 values between 4.04 and 5.161 values between 5.16 and 6.271 values between 9.62 and 10.741 values between 17.43 and 18.55offnegativepositiveoften

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. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  2. nett
  3. afen
  4. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  5. -Za
  6. %/
  7. DidLoad
  8. Hooks
  9. IRC
  10. eton

Suppresses

  1. from
  2. 从
  3. from
  4. →from
  5. 從
  6. 是从
  7. From
  8. _from
  9. From
  10. .from