Neuron Explorer

Qwen3-0.6B

unclear: mid-sentence words before a choice

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 fires on a word that sits just before a choice or a continuation, such as 'within', 'or', 'categories' or an opening bracket. When it was turned up, the text drifted into rambling, mostly unrelated wording, so no clear theme came through.

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

    smooth variations which are precise to the unique paintings.⏎Sextus Empiricus' Outlines of Pyrrhonism is among the most crucial and influential texts within the historical past of Greek philosophy. within the calls for of cause Casey Perin

  2. #2 peak

    desires, the right kind method to pursue happiness, and the correct plan for an excellent lifestyles.⏎Read or Download Aristotle for Everybody: Difficult Thought Made Easy PDF⏎Best greek &

  3. #3 peak

    total footprint of around 84 acres can be selectively opened up and intensified with vegetation.⏎Right: Microforests aim to connect existing patches of vegetation<|endoftext|>Related categories 2⏎Atmospheres of Neutron Stars⏎An interactive archive of neutron

  4. #4 peak

    to abcteach⏎Free Account (Settings)⏎An 8 page study unit for Ron Roy's A-Z Mystery,

  5. #5 peak

    opportunity to develop into the individual he or she is destined to be by virtue of their own efforts.⏎Pure Fun - Yoga

  6. #6 peak

    critical areas: (1) developing understanding of and applying proportional relationships; (2) developing understanding of operations with rational numbers and working with expressions and linear equations; (3) solving problems involving scale drawings and informal geometric constructions, and working. K

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: The method of the short-tangent is a method that is part of the algebra that is used to the method of determining the tangent to a function at the (e) the instant of the it is the method of determining the tangent at the

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

    , it should let people know the importance of saving energy. Second, efforts should be made to develop new energy resources, such as solar energy, geothermal(地热) energy and so on. We have reason to be optimistic in this respect

  2. #2 peak

    in the Nihon Shoki 日本書紀 (720) 9 and in the Hizen-no-kuni fudoki 肥前國風土記 (first part of the 8th century) 10

  3. #3 peak

    . Because terrain shadowing effects are not considered, areas with high camera angles may also appear blurred if they are located (viewed from the camera position) behind mountain ridges or peaks.⏎i.photo.rectify can be run directly, specifying options

  4. #4 peak

    the archipelago’s land surface. The fold in the Earth’s crust that formed the Kornati mountain range was produced by the collision of the Adriatic and European tectonic plates

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. L19:1118 excites
  2. L17:257 weakly suppresses
  3. L19:53 weakly suppresses
  4. L18:2998 weakly excites
  5. L16:461 weakly excites
  6. L14:392 weakly excites
  7. L0:0 weakly excites
  8. L1:2596 weakly suppresses
  9. L7:547 weakly suppresses
  10. L18:569 weakly excites
  11. L18:533 weakly excites
  12. L10:485 weakly suppresses
  13. L7:732 weakly excites
  14. L3:707 weakly suppresses
  15. L2:2036 weakly suppresses
  16. L8:328 weakly excites
  17. L13:91 weakly excites
  18. L15:313 weakly suppresses decimal point inside numbers
  19. L17:1652 weakly excites
  20. L6:8 weakly suppresses
  21. L3:421 weakly suppresses
  22. L10:157 weakly suppresses end of document and web boilerplate
  23. L20:417 weakly suppresses
  24. L16:296 weakly suppresses

Feeds later layers

  1. L22:2618 excites
  2. L23:471 weakly excites
  3. L24:766 weakly excites
  4. L27:2726 weakly excites
  5. L23:892 weakly excites "the" in "over the past" time spans
  6. L25:1263 weakly excites
  7. L22:2272 weakly excites end of text token before new document
  8. L22:2745 weakly excites
  9. L25:269 weakly excites
  10. L23:577 weakly suppresses
  11. L24:2563 weakly suppresses
  12. L24:2240 weakly suppresses
  13. L26:696 weakly suppresses
  14. L26:2055 weakly suppresses
  15. L23:2063 weakly suppresses
  16. L22:1887 weakly excites
  17. L22:1422 weakly excites
  18. L23:2671 weakly excites
  19. L25:117 weakly excites
  20. L22:1387 weakly excites
  21. L24:2718 weakly suppresses
  22. L25:2486 weakly excites
  23. L24:2009 weakly excites
  24. L25:1234 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 -8.45 and -8.201 values between -5.79 and -5.559 values between -0.97 and -0.7220 values between -0.72 and -0.4890 values between -0.48 and -0.241639 values between -0.24 and -0.002000 values between -0.00 and 0.24101 values between 0.24 and 0.4828 values between 0.48 and 0.7214 values between 0.72 and 0.975 values between 0.97 and 1.21offnegativepositiveoften

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. 区
  2. .default
  3. 区管委会
  4. 县
  5. bled
  6. IPH
  7. default
  8. íveis
  9. иона
  10. //#

Suppresses

  1. ...'
  2. ệu
  3. chnitt
  4. -quarters
  5. ')'
  6. inance
  7. ETO
  8. @Api
  9. fuss
  10. .boost