Neuron Explorer

Qwen3-0.6B

words that set up a following clause

A neuron in layer 19 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 small connecting words, like 'but', 'and', 'of' or 'missing', that lead into the next part of a sentence. It goes quiet on 'which' when a question is being introduced. Turning it up made the model's text wander into number lists and unrelated topics.

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

    than 15 PSIG but less than or equal to 125 PSIG.⏎- Regulator Size not applicable. The resulting Amount of Steam exceeds a Regulator Size of 4 inches.<|endoftext|>By: Dr. David Barker,

  2. #2 peak

    to make sure they have everything they need for the day. If I don’t triple check every detail, they’re likely to be fully prepared for snack time but missing important papers or sports equipment. What’s the right thing to do?⏎Your desire

  3. #3 peak

    non-significant LST temperature differences, result on which data source for specific time period had higher maximum LST can be mixed. This could be because of already mentioned approximate land type detection. In<|endoftext|>Nutrition plays a very important role in the

  4. #4 peak

    led the first wave, exploiting the first moments of surprise to attack the most important ships present (the battleships), while dive bombers attacked U.S.<|endoftext|>Definition of coal mine in English:⏎An excavation in the earth for extracting coal.⏎- France

  5. #5 peak

    study further concluded that deaths in the year of the study publication would be in excess of 300,000. So why are pollution levels dropping but pollution-related fatalities rising? Well, the answer to that is complex and not fully

  6. #6 peak

    that these rots continue to be an emerging threat in the PNW and will require future management attention.⏎For more information about Phacidiopycnis rot and Speck rot and their management visit our page

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 question of a business having a "90% error" to the decision made by the decision (the decision itself, which is in the question as 1000000000000000

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

    os merge at later epochs. These points will be addressed later.⏎As in the case of stellar collapses, the basic concept is Jeans' Mass. We must know which masses are able to collapse and how the collapse grows as a function of time.

  2. #2 peak

    and the intestine, plays a similar role in counteracting bacterial infections in the gastrointestinal system.⏎“We wanted to determine which sugar chain of GP2 corresponded to the high-mannose chain of U

  3. #3 peak

    as to show why it is imperative to go to a hearing specialist. Ear probes as well as visual mapping of the ear’s abilities are two ways of finding out which sounds are entering into the ear and being interpreted by the device. With this knowledge

  4. #4 peak

    being.⏎What is blood cholesterol? Also referred to simply as cholesterol, blood cholesterol refers to how much of a lipid (fat) called cholesterol we have per deciliter of blood

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. L17:1638 suppresses
  2. L18:1650 suppresses
  3. L18:1128 suppresses
  4. L17:1592 excites
  5. L17:849 suppresses
  6. L18:606 suppresses
  7. L14:2142 suppresses
  8. L15:8 excites
  9. L13:700 excites
  10. L18:95 excites end of a document marker
  11. L18:1264 excites
  12. L14:2235 weakly suppresses
  13. L17:252 weakly excites
  14. L16:896 weakly excites
  15. L14:2110 weakly excites "for" in same-or-true comparisons
  16. L18:675 weakly suppresses
  17. L14:2156 weakly suppresses opening parenthesis after a term
  18. L13:1651 weakly suppresses
  19. L18:557 weakly suppresses
  20. L18:786 weakly suppresses
  21. L18:3026 weakly suppresses
  22. L18:134 weakly suppresses
  23. L18:69 weakly excites
  24. L8:2262 weakly excites

Feeds later layers

  1. L20:1174 excites
  2. L20:25 suppresses
  3. L20:1189 excites
  4. L20:158 suppresses
  5. L20:1153 excites
  6. L20:379 suppresses digits inside dates
  7. L21:115 excites
  8. L21:105 suppresses
  9. L21:92 suppresses comma in a list or after a noun phrase
  10. L20:113 weakly suppresses
  11. L20:759 weakly excites
  12. L21:377 weakly suppresses
  13. L20:669 weakly suppresses
  14. L21:224 weakly excites
  15. L20:658 weakly suppresses
  16. L20:782 weakly excites
  17. L20:76 weakly excites
  18. L22:345 weakly excites
  19. L23:2536 weakly suppresses
  20. L27:2272 weakly suppresses
  21. L21:167 weakly suppresses
  22. L22:107 weakly excites
  23. L25:770 weakly suppresses
  24. L20:170 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 -8.16 and -7.792 values between -7.79 and -7.421 values between -7.42 and -7.056 values between -7.05 and -6.695 values between -6.69 and -6.3210 values between -6.32 and -5.954 values between -5.95 and -5.5811 values between -5.58 and -5.2219 values between -5.22 and -4.8519 values between -4.85 and -4.4841 values between -4.48 and -4.1144 values between -4.11 and -3.7545 values between -3.75 and -3.3880 values between -3.38 and -3.0177 values between -3.01 and -2.6499 values between -2.64 and -2.28145 values between -2.28 and -1.91187 values between -1.91 and -1.54237 values between -1.54 and -1.17301 values between -1.17 and -0.81356 values between -0.81 and -0.44461 values between -0.44 and -0.07588 values between -0.07 and 0.30300 values between 0.30 and 0.66237 values between 0.66 and 1.03180 values between 1.03 and 1.40139 values between 1.40 and 1.7795 values between 1.77 and 2.1373 values between 2.13 and 2.5033 values between 2.50 and 2.8739 values between 2.87 and 3.2426 values between 3.24 and 3.6018 values between 3.60 and 3.9716 values between 3.97 and 4.346 values between 4.34 and 4.714 values between 4.71 and 5.071 values between 5.07 and 5.441 values between 5.44 and 5.811 values between 6.18 and 6.54offnegativepositiveoften

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. みたいです
  2. ivamente
  3. Strat
  4. enne
  5. laughs
  6. idan
  7. 回事
  8. .optString
  9. 하겠습니다
  10. 家伙

Suppresses

  1. rink
  2. recovery
  3. diesel
  4. Infinite
  5. infinite
  6. rer
  7. Atoms
  8. Recovery
  9. /configuration
  10. autob