Neuron Explorer

Qwen3-0.6B

word-start 'le' and 'log' pieces

A neuron in layer 26 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 the short pieces 'le' and 'log' when they start a word, as in 'leopards', 'leach' or 'log data'. Turned up, the model repeated words like 'company', which does not match the trigger; turned down, it drifted toward 'log' and 'loggia'.

Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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

    the system. Incorrect data sometimes produce an erroneous result that reduces the quality of data. To handle this problem web Miner's has introduced several preprocessing methods to reduce the log data size and to eliminate incorrect records, which lead to improved quality, quantity and

  2. #2 peak

    males mate with most of the females leaving most males unsuccessful in reproduction. Unsuccessful males form small bachelor groups. Muscovy ducks have been described as having a lek mating system. (Agirrre and University of Southern Mississippi,

  3. #3 peak

    they mistake bright-colored pieces for food. Sea turtles and migrating birds can become entangled in abandoned plastic fishing nets known as “ghost nets.” Plastics also can leach harmful chemicals into the water which can be carried along the food chain back to

  4. #4 peak

    and Management Area, Belize⏎Peter Kovarik, John Shuey, and Chris Carlton⏎During the mid-late 1990's, lepidopterist John Shuey became interested in testing the widely touted concept

  5. #5 peak

    program is coded in such a way that it will never start another process. However, this is a limitation of that programm and not the operating system.⏎When you log onto a UNIX system, you usually get access to a command line interpreter, or

  6. #6 peak

    .3 kg) compared to soy (1.8 kg).⏎The U.S. Dairy Export Council proposes that whey’s naturally high levels of the amino acid leucine-an amino acid involved in the building of lean muscle-may have been

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: though loaves of loaves are loaves of loaves to loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves loaves lo

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

    olfactory system, attacking other nerve cells, and spreading to other areas of the brain.⏎In diabetic monkeys, the virus was also found in the so-called entorhinal cortex, which also plays an important role in Alzheimer’s disease. If

  2. #2 peak

    vère had set up an independent government on Providencia Island, in the western Caribbean, and Joseph was embarking on a political career of his own in Port-au-Prince, Haiti. Though the two brothers were born in the French colony of

  3. #3 peak

    muscle tension, difficulty sleeping and difficulty concentrating.⏎Obsessive Compulsive Disorder (OCD)⏎Obs<|endoftext|>A mobile phone user outside a mobile payment facility in Port-au-Prince, Haiti.⏎our goal: to alleviate poverty

  4. #4 peak

    HMCS La Malbaie⏎HMCS La Malbaie was a Royal Canadian Navy revised Flower-class corvette which took part in convoy escort duties during the Second World War. She fought

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. L23:1084 suppresses
  2. L23:166 excites
  3. L24:202 suppresses
  4. L20:97 weakly suppresses
  5. L20:110 weakly excites
  6. L18:719 weakly excites
  7. L16:91 weakly excites
  8. L23:1570 weakly excites
  9. L9:1104 weakly suppresses the verb 'visit' before a website
  10. L24:2157 weakly excites
  11. L16:484 weakly suppresses
  12. L25:215 weakly excites
  13. L10:1521 weakly excites
  14. L22:1433 weakly excites
  15. L2:2889 weakly suppresses
  16. L1:1602 weakly suppresses
  17. L17:542 weakly suppresses
  18. L23:2482 weakly excites
  19. L3:947 weakly excites
  20. L3:1181 weakly excites
  21. L3:763 weakly suppresses
  22. L1:983 weakly excites
  23. L6:1091 weakly suppresses
  24. L10:276 weakly excites

Feeds later layers

  1. L27:827 strongly suppresses
  2. L27:2824 weakly suppresses
  3. L27:1940 weakly excites
  4. L27:943 weakly suppresses
  5. L27:2268 weakly excites
  6. L27:2240 weakly suppresses
  7. L27:1898 weakly suppresses
  8. L27:2907 weakly suppresses the letter p starting a word piece
  9. L27:1805 weakly excites
  10. L27:1786 weakly suppresses
  11. L27:120 weakly excites
  12. L27:1851 weakly excites capital K starting a word
  13. L27:1765 weakly suppresses
  14. L27:3059 weakly excites
  15. L27:1590 weakly excites
  16. L27:1084 weakly suppresses
  17. L27:2874 weakly excites
  18. L27:676 weakly excites
  19. L27:941 weakly excites
  20. L27:1989 weakly excites
  21. L27:1334 weakly excites
  22. L27:473 weakly excites
  23. L27:2007 weakly suppresses
  24. L27:1394 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 -13.90 and -12.235 values between -10.56 and -8.891 values between -8.89 and -7.221 values between -7.22 and -5.566 values between -5.56 and -3.8910 values between -3.89 and -2.22164 values between -2.22 and -0.553680 values between -0.55 and 1.1230 values between 1.12 and 2.795 values between 2.79 and 4.461 values between 4.46 and 6.122 values between 6.12 and 7.791 values between 7.79 and 9.461 values between 51.18 and 52.84offnegativepositiveoften

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. فتر
  2. SharedPtr
  3. entrant
  4. Angebot
  5. 友善
  6. ely
  7. ements
  8. .appcompat
  9. 主体
  10. anteed

Suppresses

  1. log
  2. log
  3. Log
  4. (log
  5. logs
  6. Log
  7. $log
  8. =log
  9. →log
  10. logs