Neuron Explorer

Qwen3-0.6B

'back' after 'dates' or 'dating'

A neuron in layer 3 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 'back' (and 'to') when it follows 'dates' or 'dating' in phrases like 'dates back to' or 'dating back thousands of years'. When turned up, the model's writing did not obviously change in this way.

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

    largest manmade prehistoric mound, Silbury Hill.⏎Avebury's stone circle dates to around 2600 B.C. Experts are still trying to determine its

  2. #2 peak

    certain nanotechnological applications as well, including attempts to create more durable and sustainable materials and substances for manufacturing industries.⏎The use of computer technology in engineering fields dates back to the 80s, but recent years have seen the two become almost

  3. #3 peak

    very difficult. The products of CO2 partial reduction, formic acid and formates, are likewise resistant to reduction.1 CO2 electro-reduction studies date back to experiments done on zinc more than a century ago.2, 3 In

  4. #4 peak

    offices. Most states have laws regulating the operation of lotteries, and some even require retailers to be licensed before selling them.⏎The history of the lottery dates back hundreds of years, with some of the first known lotteries being held in

  5. #5 peak

    advice. Please talk to your doctor before using any of the herbs and/or remedies mentioned in this article. Yarrow is one of the oldest medicinal herbs, dating back thousands of years and used for various ailments. It’s a member of the aster

  6. #6 peak

    a peal of bells, swing paper balls at each other and scramble for pennies. Up-Helly-Aa is a Shetland Isles festival, dating back to Viking times, when a thirty-foot model Viking ship, complete with banners,

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: that the number of people who are in a car is a small thing. So, I think that a person who has a certain number of people who are in a car. Maybe, I don't know. I think that the number of

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

    densities to make our diagnosis. Our historic false negative result of 17.<|endoftext|>Native Animals of the Thar Desert⏎The Thar Desert, the pride of Rajasthan, is also called an ocean of sand. The Thar or Great Indian

  2. #2 peak

    kids was aiming for its broad, flat leaves, but hit the stem instead, leaving a clean, gaping hole, from one side to the other.⏎William Wcislo, an entomologist at the Smithsonian Tropical Research Institute and father of two

  3. #3 peak

    Azteca alfari, small, sand-colored ants known to build nests inside cecropia trees, which are widespread in Central and South America. Wcislo saw how they swarmed to the site of injury, gathered small, fib

  4. #4 peak

    the state;⏎ii) all states of India to be equally weighted;⏎iii) the UT of Chandigarh, and Punjabi speaking areas which were 'deliberately left out' post-1966 trifurcation to be transferred

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. L1:1801 weakly excites
  2. L2:70 weakly excites
  3. L2:489 weakly suppresses
  4. L2:1590 weakly suppresses
  5. L2:980 weakly excites
  6. L2:1367 weakly excites
  7. L1:1205 weakly suppresses
  8. L1:2683 weakly excites
  9. L2:1036 weakly excites
  10. L0:2295 weakly excites
  11. L1:2922 weakly excites
  12. L2:75 weakly suppresses
  13. L0:516 weakly excites
  14. L1:2677 weakly suppresses
  15. L1:1340 weakly suppresses
  16. L1:146 weakly excites
  17. L0:535 weakly excites
  18. L0:2269 weakly suppresses
  19. L0:371 weakly suppresses
  20. L0:1251 weakly excites
  21. L1:1152 weakly excites
  22. L2:1687 weakly suppresses
  23. L2:2190 weakly excites
  24. L2:2068 weakly suppresses

Feeds later layers

  1. L9:1639 weakly excites
  2. L4:1509 weakly suppresses
  3. L22:2813 weakly excites
  4. L12:717 weakly excites
  5. L17:2887 weakly excites
  6. L13:875 weakly suppresses
  7. L6:778 weakly excites
  8. L7:270 weakly excites
  9. L18:1165 weakly excites
  10. L11:2013 weakly excites
  11. L7:2572 weakly suppresses
  12. L23:2962 weakly excites
  13. L23:2853 weakly excites
  14. L22:2238 weakly suppresses
  15. L22:1152 weakly excites
  16. L20:2675 weakly suppresses
  17. L5:833 weakly excites
  18. L7:962 weakly excites
  19. L5:459 weakly excites
  20. L12:434 weakly suppresses
  21. L18:1789 weakly suppresses
  22. L11:1004 weakly excites
  23. L15:614 weakly excites
  24. L23:2245 weakly excites "at" before end, bottom or start

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 -2.93 and -2.741 values between -2.18 and -1.991 values between -1.24 and -1.061 values between -0.87 and -0.685 values between -0.68 and -0.4931 values between -0.49 and -0.31385 values between -0.31 and -0.122290 values between -0.12 and 0.071083 values between 0.07 and 0.2695 values between 0.26 and 0.447 values between 0.44 and 0.632 values between 0.63 and 0.821 values between 0.82 and 1.001 values between 1.00 and 1.191 values between 1.19 and 1.381 values between 1.75 and 1.941 values between 2.88 and 3.061 values between 4.37 and 4.56offnegativepositiveoften

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. 是一部
  2. Pediatrics
  3. innacle
  4. .jquery
  5. acam
  6. asm
  7. debugger
  8. ean
  9. legacy
  10. aurus

Suppresses

  1. nicos
  2. icopt
  3. 媒
  4. bable
  5. 强度
  6. held
  7. he
  8. vetica
  9. radio
  10. RADIO