Neuron Explorer

Qwen3-0.6B

unclear: word fragments across languages

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 partial word pieces in many languages, such as Russian, Finnish and Persian, often in the middle of a word. Its pattern is not clear, and turning it up produced garbled, repeated text.

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

    евой колонке, с их английскими эквивалентами, данными в правой колонке⏎1) legislative branch a) Член Верховного Суда⏎2) civil and criminal cases b)

  2. #2 peak

    techniques and processes that were developed *before* the idea of genetic engineering which is the biochemical creation of specific advantageous mutations in the genome of a plant. Before the advent of GMO farmers and plant breeders had to wait for nature to create a natural

  3. #3 peak

    yväislinkit|⏎|Tila||Julkaistu - 1 syysk. 2021|⏎|OKM-julkaisutyyppi||A1 Julkaistu artikkeli, so

  4. #4 peak

    it. What this implies is that caffeine is doing something to your internal brain chemistry to make this happen – and where there’s neurochemistry, there’s a ton of crazy internal mechanisms that actually make that effect happen.⏎One can easily and efficiently

  5. #5 peak

    and the Zoological Society of London’s Institute of Zoology and published this week in the journal Nature Communications⏎The African straw-coloured fruit bat (Eidolon helv

  6. #6 peak

    garrot, gowdy, pied widgeon, whiteside, curre, and doucker. Barrow's golden-eye of America (Glaucionetta Islandica) is

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: there is a one-of-100-100-100 There is a one-of-100-100-100 probabilities that there is either of them, either of them, or

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

    (i.e. Varro's description of the extent of Roma quadrata) mit dem unten erörterten Pomerium und der Auffassung des Dionys. I 88: περιγράφ

  2. #2 peak

    Antalya-Türkiye’de doğdu. Lisans öğrenimini MSÜ Deniz Harp Okulunda tamamlamıştır.⏎Yüksek Lisans Öğrenim

  3. #3 peak

    ментарии при наличии в них нецензурной брани, оскорблений или спама<|endoftext|>Conflict and Solid

  4. #4 peak

    их английскими эквивалентами, данными в правой колонке⏎1) legislative branch a) Член Верховного Суда⏎2) civil and criminal cases b) окружной суд⏎3

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. L18:893 excites
  2. L16:864 excites
  3. L18:2342 excites
  4. L18:2676 excites
  5. L20:2873 suppresses
  6. L19:939 suppresses
  7. L18:550 suppresses
  8. L19:2507 excites
  9. L17:500 suppresses
  10. L18:491 excites
  11. L16:1231 excites
  12. L7:122 excites
  13. L7:76 excites
  14. L20:2800 suppresses
  15. L20:789 excites
  16. L20:1796 excites
  17. L19:2880 excites
  18. L19:1025 suppresses
  19. L20:1356 suppresses
  20. L19:1584 excites
  21. L19:1581 suppresses
  22. L19:2640 weakly excites
  23. L19:3044 weakly excites
  24. L20:677 weakly excites

Feeds later layers

  1. L22:2745 strongly excites
  2. L27:2096 strongly suppresses
  3. L24:359 strongly suppresses
  4. L26:1320 suppresses
  5. L25:2973 suppresses
  6. L23:1390 excites
  7. L27:1582 excites
  8. L23:258 suppresses
  9. L23:2526 excites
  10. L27:2605 excites
  11. L25:284 suppresses
  12. L26:711 suppresses 'years' after an age number
  13. L22:142 excites 'of' after a quantity or noun phrase
  14. L22:171 excites
  15. L27:1260 weakly suppresses
  16. L26:2643 weakly excites
  17. L27:1737 weakly suppresses
  18. L22:1425 weakly excites
  19. L22:1523 weakly suppresses
  20. L22:309 weakly excites
  21. L27:1764 weakly suppresses
  22. L22:1993 weakly suppresses
  23. L25:917 weakly suppresses
  24. L26:507 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 -3.67 and -3.441 values between -3.20 and -2.961 values between -2.72 and -2.481 values between -1.76 and -1.531 values between -1.29 and -1.051 values between -1.05 and -0.8112 values between -0.81 and -0.5742 values between -0.57 and -0.33192 values between -0.33 and -0.092714 values between -0.09 and 0.15644 values between 0.15 and 0.38178 values between 0.38 and 0.6275 values between 0.62 and 0.8625 values between 0.86 and 1.1010 values between 1.10 and 1.341 values between 1.58 and 1.822 values between 1.82 and 2.063 values between 2.29 and 2.531 values between 2.77 and 3.011 values between 3.49 and 3.731 values between 3.73 and 3.971 values between 5.64 and 5.88offnegativepositiveoften

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. يون
  2. ость
  3. voie
  4. ные
  5. reek
  6. такие
  7. деятельность
  8. تان
  9. жение
  10. różne

Suppresses

  1. ального
  2. енного
  3. ului
  4. ического
  5. arters
  6. ителей
  7. емых
  8. ющей
  9. альных
  10. телей