end-of-text token mid-document
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 18 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron reacts to the hidden 'end of document' marker that the training data uses to join unrelated web pages. It is most active right where one page seems to stop and another starts. Turned up, it made the model write repeated filler, and turned down it made the text drift toward puzzles and quiz-style questions.
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 activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
to look up "vocable"? Please share what surprised you most...<|endoftext|>Biggles Sees It Through⏎|Published||Aug-41 - Oxford
-
arnessed this sequence to overcome a myriad of hurdles. plants use it to sprout branches no one gets covered and each gets maximum amount of sunlight, and as<|endoftext|>Помогите пожаааалуйста, сам не с
-
413/Healthy Hunger Free Kids Act implementation; (2) assess changes in the proportion of schools with excellent water access and barriers and facilitators to More<|endoftext|>Minnesota is a disclose in the Upper Midwest,
-
two different numbers cant have the same cube). Therefore, by Definition 1, is one-to-one....⏎View Full Document⏎- Spring '09<|endoftext|>part of speech: noun⏎definition 1: an upright line or pole with
-
by 5 and 7, then increase⏎count by one.⏎Share your code⏎Show a friend, family member, or teacher what you've done!<|endoftext|>cigarette butts are the most common form of plastic litter
-
puzzle.⏎Click for Sudoku Puzzles from previous weeks.⏎Click for tips for teaching students to do Sudoku puzzles.⏎Copyright © 2006 Education World<|endoftext|>1.Modeling design and nozzle selection of musical fountain .⏎For the overall design of
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
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.
Here is something I have been thinking about lately: the **BisQUO**. I have a lot of questions that I want to ask, but I think they need to be raised in the **2000**. I am aware of the 2000,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the synergy of every team member.⏎n.大綱, 提要 [ summary or outline of a book, play, etc ]⏎note: 字
-
tornado one day? Michaud’s prototypes have all proved successful but no investors have expressed an interest as of yet.⏎Dissolving the Dead⏎Traditional burials raise a
-
:⏎Reaching a wider audience effectively! : Because an instructor can handle a small mob at a time only. When every other student bombards him with questions and raises hand in every micro second, the teacher sits down with head in his hands.⏎
-
). It is one of Indonesia's most active volcanoes and lies in one of the world's most densely populated areas. Many people raise crops and livestock in its shadow.⏎Yudhoyono has announced that residents
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L17:700 weakly excites
- L6:16 weakly excites
- L2:1499 weakly suppresses
- L3:1702 weakly excites
- L13:2324 weakly suppresses
- L17:1445 weakly excites
- L10:1910 weakly suppresses
- L13:118 weakly suppresses
- L5:1722 weakly suppresses
- L16:193 weakly suppresses 'wide' in 'a wide range of'
- L2:75 weakly suppresses
- L16:153 weakly excites
- L3:278 weakly excites
- L16:2578 weakly excites
- L6:17 weakly suppresses
- L16:1156 weakly excites
- L6:566 weakly suppresses
- L11:2304 weakly suppresses
- L3:1075 weakly suppresses
- L5:850 weakly suppresses
- L16:2785 weakly suppresses
- L6:1998 weakly suppresses
- L10:1795 weakly suppresses
- L5:701 weakly excites
Feeds later layers
- L20:1127 weakly suppresses
- L25:2603 weakly excites
- L19:996 weakly suppresses the digit 5 inside numbers
- L22:1915 weakly excites
- L22:1432 weakly suppresses
- L20:466 weakly excites
- L22:2913 weakly suppresses
- L19:2592 weakly excites
- L25:2650 weakly suppresses
- L22:1100 weakly suppresses
- L19:1810 weakly excites
- L27:1473 weakly suppresses
- L27:391 weakly suppresses
- L19:3063 weakly excites
- L20:1669 weakly suppresses
- L21:814 weakly suppresses
- L24:2168 weakly suppresses
- L19:623 weakly suppresses
- L20:1964 weakly suppresses
- L19:2568 weakly suppresses
- L23:2348 weakly suppresses
- L23:1396 weakly excites
- L21:604 weakly suppresses
- L21:9 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
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.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
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 layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- 徂
- moth
- tempts
- 嫖
- letal
- tempt
- casts
- ASM
- unnel
- ographical
Suppresses
- 异议
- ayar
- 意见
- enor
- 醒
- ands
- ęb
- ило
- 高手
- 捧