Neuron Explorer

Qwen3-0.6B

unclear: pieces of split words

A neuron in layer 4 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 weakly on the middle pieces of words and names that the model splits into fragments, such as parts of 'Kibo', 'Koss' or 'Cossack'. It is weakly negative on fragments like 'circum' and 'vigil'. Turning it up or down did not change the writing in a clear way.

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

    refers to the ability of a bank or banks to meet the demands of its depositors. As the banking system is highly gicd, it does not have on hand all of the cash deposited in its system

  2. #2 peak

    peak in Africa. Mount Kilimanjaro is not just an ordinary mountain, it is the world’s largest stratovolcano with three volcanic cones: Kibo, Mawenzi and Shira. Though it appears to be dormant,

  3. #3 peak

    gap significantly.⏎Viennese researchers have developed a new route for the first truly bulk production of LLCCs, i.e. carbyne, confined inside DWCNTs. From our Raman and TEM studies with contemporary ab-initio calculations

  4. #4 peak

    was 15.2 per cent in 2006. This shows a commendable reduction of poverty. The Department of Census and Statistics (DCS) declares that the OPL is Rs. 3,7

  5. #5 peak

    well known are the Polish Cossacks (Kozacy) and the Tatar Cossacks (Nağaybäklär). The name “Cossack” should not be confused with Kazakhs (native to Kazakhstan, a

  6. #6 peak

    should be.⏎The survey was conducted during May to July 2006, and towards the end of that period France lost the World Cup final after Zinedine Zidane was sent off for headbutting an Italian player. Despite

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: , I have been thinking about lately: , I have been thinking about lately: , I have been thinking about lately: , I have been thinking about lately: , I have been thinking about lately

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

    collective—public—process of value creation, and enabled more value extraction. Yet these stories have become accepted as common sense—always a term to be treated circumspectly. We have become accustomed to much talk about the pros and cons of

  2. #2 peak

    ).⏎Values can be taught through a dedicated curriculum. That is why my study is situated in<|endoftext|>A Cancer charity advices to watch out for symptoms.⏎Constant Vigilance for cancer recurrence has its own downsides.⏎After successful cancer treatment, being

  3. #3 peak

    exposure to death challenges the very premise of warfare.⏎The global ‘War on Terror’ has ushered in a new kind of warfare. It is no longer circumscribed in time or location, but has been extended indefinitely, operating transnationally

  4. #4 peak

    the disease: swelling of twigs and roots, yellowing of leaves and deformation of organs.⏎The fight against the “Swollen shoot” currently consists of circumscribing it, since there is to date no cure for this virus . This process

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. L3:27 suppresses
  2. L2:1599 suppresses
  3. L3:256 suppresses
  4. L3:195 suppresses
  5. L3:28 suppresses
  6. L2:1816 excites
  7. L3:211 suppresses
  8. L1:119 excites
  9. L2:115 suppresses
  10. L3:786 suppresses
  11. L3:264 excites
  12. L1:1097 suppresses
  13. L3:1046 excites
  14. L0:1075 suppresses
  15. L2:1464 weakly suppresses
  16. L3:303 weakly suppresses
  17. L1:861 weakly excites
  18. L0:639 weakly suppresses
  19. L2:1628 weakly suppresses
  20. L1:678 weakly suppresses
  21. L1:84 weakly suppresses
  22. L1:1774 weakly excites
  23. L0:616 weakly suppresses
  24. L1:109 weakly excites

Feeds later layers

  1. L6:58 excites
  2. L9:708 excites
  3. L5:365 excites
  4. L9:132 suppresses
  5. L7:220 suppresses
  6. L5:230 excites
  7. L9:354 excites
  8. L15:74 suppresses
  9. L6:182 suppresses
  10. L5:14 suppresses
  11. L13:695 excites
  12. L5:315 weakly excites
  13. L9:437 weakly excites
  14. L17:798 weakly excites
  15. L9:324 weakly excites
  16. L10:1815 weakly excites
  17. L8:855 weakly excites
  18. L27:181 weakly suppresses
  19. L10:984 weakly suppresses
  20. L15:693 weakly excites commas inside numbers and asides
  21. L15:2065 weakly excites
  22. L18:2829 weakly excites
  23. L16:1596 weakly excites
  24. L12:1902 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 -1.83 and -1.771 values between -1.22 and -1.161 values between -1.16 and -1.092 values between -1.09 and -1.023 values between -1.02 and -0.953 values between -0.95 and -0.891 values between -0.89 and -0.824 values between -0.82 and -0.757 values between -0.75 and -0.688 values between -0.68 and -0.6212 values between -0.62 and -0.5518 values between -0.55 and -0.4832 values between -0.48 and -0.4136 values between -0.41 and -0.3466 values between -0.34 and -0.28109 values between -0.28 and -0.21131 values between -0.21 and -0.14346 values between -0.14 and -0.07747 values between -0.07 and -0.011054 values between -0.01 and 0.06593 values between 0.06 and 0.13352 values between 0.13 and 0.20175 values between 0.20 and 0.27109 values between 0.27 and 0.3351 values between 0.33 and 0.4014 values between 0.40 and 0.4712 values between 0.47 and 0.547 values between 0.54 and 0.604 values between 0.60 and 0.673 values between 0.67 and 0.744 values between 0.74 and 0.812 values between 0.81 and 0.88offnegativepositiveoften

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. 当地人
  2. ^-
  3. -⏎
  4. ornings
  5. ứt
  6. updatedAt
  7. WebClient
  8. editary
  9. ».
  10. Investments

Suppresses

  1. جب
  2. buie
  3. 几次
  4. BED
  5. rium
  6. ыв
  7. że
  8. aab
  9. يرة
  10. rise