commas in catalog, date, and list metadata
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. 10 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 lights up on commas in dry, structured text such as library catalog entries, publication details, dates, and short lists of tags or categories. When it was turned up, the model's writing did not show a clear theme; it drifted into repeated numbers and lists.
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.
-
al., 2001; CIA, 2003). Furthermore, parks and<|endoftext|>From: Mary Ann Liebert Inc.⏎Posted: Saturday, September 1, 2007⏎Astrobiology August 2
-
found in the catalog.⏎|Statement||World Council of Churches|⏎|Publishers||World Council of Churches|⏎|The Physical Object|⏎|Pagination||xvi, 80 p. :|⏎|Number of Pages||99|⏎|
-
Cell Biology⏎Chlamydomonas; Dynein ATPase; Flagella; Microscopy, Electron; Microtubules⏎Analysis of serial cross-sections of the Ch
-
the Pharaohs⏎World bibliography<|endoftext|>Also found in: Dictionary, Thesaurus, Medical, Wikipedia.⏎cavities connected to the respiratory pathways
-
the Pew Research Center shows that Muslim countries on average… Latest update on: 2006-SEP-23 Matthew 12:30, Luke 11:23, and Romans 16:17
-
boxes familiarize students with information needed to accurately solve drug calculation problems.⏎PUBLICATION DATE: 9/4/2009⏎CATEGORY: Mathematics, Medical, Science<|endoftext|>International solidarity: creating a world for all agesInternational solidarity:
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: I have a book manuscript that I have a copy of, of the book of the Bible in the hand of the reader of the user: : "I have a manuscript: "I have a" in the hand of the user
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
six grade-appropriate math problems: one each of addition, subtraction, multiplication, division, fractions, and percents. I stressed "grade
-
2014, thirteen animals have tested positive for rabies in Erie County (five raccoons, five bats, and one each of a cat, skunk and fox).⏎On June 11th, a 15
-
1944.⏎Others are Kolmaneskuppe, Namibia, a former diamond mining town whose population steadily declined after WWI,
-
35%, 34%, 9% and 6%, respectively, of all cases reported by EU/
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
- L5:16 strongly excites
- L4:3 strongly excites commas after citation fields
- L9:77 strongly excites commas in citations and name lists
- L2:66 excites
- L3:71 suppresses
- L0:2921 suppresses
- L6:305 excites
- L9:1113 suppresses
- L9:2447 suppresses
- L3:758 weakly excites
- L9:1675 weakly excites
- L3:487 weakly suppresses
- L5:102 weakly excites
- L4:295 weakly excites
- L5:1304 weakly excites
- L3:392 weakly suppresses
- L6:250 weakly excites
- L0:68 weakly suppresses
- L4:1867 weakly excites
- L6:655 weakly excites
- L8:441 weakly excites 'or' linking alternatives
- L9:1421 weakly suppresses
- L5:210 weakly excites
- L6:2925 weakly excites
Feeds later layers
- L14:35 strongly excites
- L16:240 strongly excites comma after 'sometimes' or 'the truth'
- L17:2001 excites
- L19:1826 excites
- L18:366 excites comma inside numbers and dates
- L17:118 excites comma after a clause ending
- L22:144 excites
- L17:1405 excites
- L20:1298 excites
- L16:2745 excites
- L11:154 excites commas after citation numbers
- L15:588 excites comma after an introductory phrase
- L19:726 excites comma inside large numbers
- L15:1213 excites
- L19:2645 excites commas in the middle of a sentence
- L18:1394 excites
- L17:627 excites
- L17:116 suppresses
- L16:1082 excites comma linking two clauses after 'and'
- L18:1270 excites
- L14:801 excites 'and' or 'or' joining list items
- L17:585 excites
- L23:184 excites
- L18:2043 excites
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
- która
- igung
- inke
- iaux
- ämp
- póź
- especially
- ʈ
- itize
- place
Suppresses
- ,
- иков
- /im
- im
- egra
- iao
- ={
- ({
- $,
- 越好