Neuron Explorer

Qwen3-0.6B

unclear: words inside technical terms

A neuron in layer 26 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

The pattern is hard to pin down. It fires on pieces of specialist terms, such as the 'acet' in thallium acetate or the 'epith' in epithet, but the examples do not share a clear theme. Turning it up made the text repeat words like 'pair' and 'call'.

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

    from its application. In 1938, “Kormelu,” a depilatory cream, was being sold with the active ingredient thallium acetate.(3) Thallium acetate is used to kill rats and was

  2. #2 peak

    and desires. Reason comprises conception, judgment, reasoning, and the intuitional faculty. Specifically, it is the intuitional faculty, or the faculty of first truths, as distinguished from the understanding

  3. #3 peak

    emergency kits for home and school.<|endoftext|>|This article needs additional citations for verification. (July 2012) (Learn how and when to remove this template message)|⏎Sendust is a magnetic metal powder that was invented by Hak

  4. #4 peak

    he was noted for his lack of military skill, and his controversial command of New Orleans, which brought him wide dislike in the South and the "Beast" epithet. He helped create the legal idea of effectively freeing fugitive slaves by design

  5. #5 peak

    example?⏎Surname is defined as the family or last name. An example of a surname is Smith when the person’s full name is John Smith. A nickname or epithet added to a person’s name.⏎What is surname in Tagalog?⏎Translation

  6. #6 peak

    Robert Winglee of the University of Washington, Seattle, calculated that some parts of the moon spend about a quarter of the moon's orbit protected by Earth's magnetosphere (Geophysical Research Letters, DOI: 10.102

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: opening location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location location

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

    the driving force behind founder and CEO Melissa Clark-Reynolds’ decision to create MiniMonos in the first place. After Melissa saw the movie ‘An Inconvenient Truth’, she decided to contact ex-US Vice President Al Gore to offer her

  2. #2 peak

    19. The massacre took place in a public park called Jallianwala Bagh in the city of Amritsar where British Indian army’s Colonel Reginald Dyer on 13 April 1919 ordered his

  3. #3 peak

    there’s a big difference: you can knock your opponents’ rocks away from the target, like the balls in croquet. Also, there are four people to a team, and a team leader called the skip, who yells instructions at the

  4. #4 peak

    most burials are carried out by members of the Ebola response team. But these measures have created a sense of mistrust between community members and health workers.⏎A Congolese psychologist and counselors at a group called the Bethesda Counselling Centre are helping

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. L20:889 excites
  2. L21:1888 suppresses
  3. L25:284 suppresses
  4. L24:153 suppresses
  5. L24:489 excites
  6. L23:79 excites
  7. L20:602 suppresses
  8. L19:1427 excites
  9. L22:260 suppresses
  10. L18:216 excites
  11. L15:184 suppresses
  12. L25:390 suppresses
  13. L20:1047 suppresses
  14. L19:1804 suppresses
  15. L19:1180 suppresses
  16. L18:1433 suppresses
  17. L21:592 excites
  18. L7:92 suppresses
  19. L22:2089 excites
  20. L21:68 excites
  21. L19:2956 weakly excites
  22. L10:365 weakly suppresses
  23. L21:338 weakly excites
  24. L22:760 weakly excites

Feeds later layers

  1. L27:478 weakly suppresses
  2. L27:2393 weakly suppresses
  3. L27:1586 weakly excites
  4. L27:2658 weakly suppresses
  5. L27:740 weakly excites
  6. L27:2732 weakly suppresses
  7. L27:2804 weakly suppresses
  8. L27:754 weakly suppresses
  9. L27:1452 weakly suppresses
  10. L27:72 weakly suppresses
  11. L27:2365 weakly excites
  12. L27:1582 weakly suppresses
  13. L27:793 weakly excites
  14. L27:1106 weakly suppresses
  15. L27:2921 weakly excites
  16. L27:2995 weakly suppresses
  17. L27:465 weakly suppresses
  18. L27:1485 weakly suppresses
  19. L27:2043 weakly suppresses
  20. L27:2985 weakly excites
  21. L27:2959 weakly suppresses
  22. L27:1908 weakly suppresses the 'd' that starts 'die' or 'dead'
  23. L27:914 weakly suppresses
  24. L27:2632 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 -22.41 and -21.041 values between -19.68 and -18.312 values between -18.31 and -16.952 values between -16.95 and -15.591 values between -15.59 and -14.225 values between -14.22 and -12.868 values between -12.86 and -11.4911 values between -11.49 and -10.1315 values between -10.13 and -8.7717 values between -8.77 and -7.4037 values between -7.40 and -6.0458 values between -6.04 and -4.6762 values between -4.67 and -3.31141 values between -3.31 and -1.95338 values between -1.95 and -0.582176 values between -0.58 and 0.78469 values between 0.78 and 2.15240 values between 2.15 and 3.51134 values between 3.51 and 4.8874 values between 4.88 and 6.2446 values between 6.24 and 7.6025 values between 7.60 and 8.9724 values between 8.97 and 10.336 values between 10.33 and 11.706 values between 11.70 and 13.062 values between 14.42 and 15.792 values between 17.15 and 18.521 values between 21.24 and 22.611 values between 23.97 and 25.341 values between 25.34 and 26.701 values between 28.06 and 29.431 values between 30.79 and 32.16offnegativepositiveoften

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. call
  2. 对
  3. Pairs
  4. pairwise
  5. rule
  6. hof
  7. ОС
  8. pairs
  9. comes
  10. AGMA

Suppresses

  1. ards
  2. ategory
  3. incarnation
  4. utos
  5. elope
  6. lland
  7. opening
  8. eler
  9. การ
  10. lero