Neuron Explorer

Qwen3-0.6B

comma before an added clause

A neuron in layer 18 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 a comma that is followed by an extra explanation, such as "cancer, the treatments chosen" or "costly, especially for". When turned up, the model's text started piling up "as well as" phrases.

Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).

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

    molecular mechanisms and biomarker identification of COVID-19 progression. However, the field progress has been hindered by the lack of robustness and reproducibility, and leaves questions with regard to the integration and functionalization of sensing elements and the

  2. #2 peak

    respond to treatment. Prognosis and survival depend on many factors. Only a doctor familiar with your medical history and the type, stage and characteristics of your cancer, the treatments chosen and the response to treatment can put all of this information together with

  3. #3 peak

    developments also played a role in Dr. Dyer's decision to write a new edition of War. Recently published findings on primates and the nature of early man, especially the behaviour of hunter-gatherer societies, have caused him to change some

  4. #4 peak

    or inward rolling of the foot), and running style, which can be influenced by strengthening the hip and core muscles.⏎A runner who increases mileage or intensity too quickly, such as by doing too many speed workouts or overracing, can develop shin

  5. #5 peak

    further costs.⏎4. Challenges and Considerations:⏎a. Affordability: Medical insurance premiums can be costly, especially for those without employer-sponsored coverage. Government subsidies and tax credits can help make

  6. #6 peak

    to wipe the fridge.⏎· Avoid using any soap or detergent, as they might leave behind a residue or scent, which can contaminate your food.⏎4. Do not store expired food products

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: "In, for, and, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to, to

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

    generation consist of the electron (e−⏎) and electron-neutrino (νe⏎). The second generation consist of the muon (μ−⏎) and muon neutrino (νμ⏎) The third generation consist of the

  2. #2 peak

    01, the European Society of Cardiology and the North American Society of Pacing and Electrophysiology proposed a classification that takes into consideration both anatomic features and electrophysiologic mechanisms determined at the time of electrophysiologic testing

  3. #3 peak

    g. Anything below 90 mm Hg is considered as low blood pressure. Generally, low blood pressure is characterized by a decrease in both the systolic as well as diastolic blood pressure.⏎• However, it may happen sometimes that

  4. #4 peak

    clients should perform correctly without them, and so many people omit them. Similarly the header of a document (the bit containing the title) should begin with <HEAD> and end with </HEAD> and the body of a document should begin

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. L16:1119 suppresses
  2. L15:346 excites
  3. L17:632 excites unclear: coordinating words in lists
  4. L16:1766 excites
  5. L16:774 suppresses
  6. L15:732 excites
  7. L17:2529 suppresses
  8. L16:473 excites
  9. L16:2360 excites
  10. L15:1339 suppresses
  11. L15:2271 excites
  12. L15:2050 excites
  13. L15:1467 excites
  14. L14:2341 suppresses
  15. L17:160 excites
  16. L17:2155 excites
  17. L16:2183 excites
  18. L15:2619 excites
  19. L14:1967 excites
  20. L16:1141 excites
  21. L14:2880 suppresses
  22. L15:1960 suppresses
  23. L15:389 suppresses
  24. L17:2911 suppresses

Feeds later layers

  1. L20:1009 strongly excites
  2. L19:2416 excites
  3. L19:164 suppresses
  4. L19:1570 excites
  5. L20:610 excites
  6. L19:2493 excites
  7. L19:290 suppresses
  8. L19:492 excites
  9. L24:205 excites
  10. L22:1663 excites
  11. L22:632 excites
  12. L19:1605 excites
  13. L22:368 excites
  14. L20:1758 excites
  15. L19:1414 excites
  16. L25:6 suppresses
  17. L22:158 suppresses
  18. L25:2003 excites
  19. L20:373 excites
  20. L27:289 excites
  21. L22:1295 excites 'as' before 'well' or 'as much'
  22. L19:979 weakly excites
  23. L22:649 weakly excites
  24. L20:2392 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 -5.19 and -4.951 values between -2.84 and -2.612 values between -2.61 and -2.372 values between -2.14 and -1.914 values between -1.91 and -1.671 values between -1.67 and -1.446 values between -1.44 and -1.2015 values between -1.20 and -0.9725 values between -0.97 and -0.7390 values between -0.73 and -0.50346 values between -0.50 and -0.261152 values between -0.26 and -0.031517 values between -0.03 and 0.21498 values between 0.21 and 0.44117 values between 0.44 and 0.6737 values between 0.67 and 0.9125 values between 0.91 and 1.1420 values between 1.14 and 1.389 values between 1.38 and 1.617 values between 1.61 and 1.8510 values between 1.85 and 2.084 values between 2.08 and 2.323 values between 2.32 and 2.553 values between 2.55 and 2.792 values between 2.79 and 3.025 values between 3.02 and 3.253 values between 3.25 and 3.491 values between 3.49 and 3.722 values between 3.96 and 4.19offnegativepositiveoften

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. 以及
  2. 以及其他
  3. 并且
  4. ]|
  5. 及
  6. ]+\
  7. 或者是
  8. 同时也
  9. plus
  10. 同时还

Suppresses

  1. embers
  2. .defaultValue
  3. .Runtime
  4. 城县
  5. ────────
  6. Beth
  7. Dresden
  8. JsonProperty
  9. artin
  10. .basicConfig