blank before a number in citations
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. 5 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 gap just before a number in legal citations, dates and verse references, such as "§ 1395" or "March 3, 1879". Its negative side appears in hyphenated or underscored words. Turning it up gave unrelated text, so the steering does not confirm the pattern.
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.
-
is Section 1185 of Title 23 of the Vermont Statutes Annotated. And the federal citation 42 U.S.C. § 1395 tells us that this cited federal statute can be found in Title
-
be readily restored to shoot, automatically more than one shot, without manual reloading, by a single function of the trigger.” 26 U.S.C. § 5845(b). The machine gun was first widely used during World War
-
history, the U.S. Civil War, was resolved by amendments—creating the legal foundation for the equal rights of African Americans. The 13th, 14th and 15h Amendments—collectively known as the Civil
-
the Lord for his unfailing love and his wonderful deeds for mankind” (107:1, 2, 8, New International Version).⏎This is a recognition that God,
-
Irene and Tropical Storm Lee,<|endoftext|>Back to: Lesson 40⏎Go on to: Lesson 42⏎Key Points: Pieces 83 & 84⏎Pieces 83 is a great little tune but difficult. Pay
-
second director). On March 3, 1879, the 45th Congress and President Hayes agreed to
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) - but I can't, so? - This is a question about a question, but it's not just a question. (the) - This is a question about a question, but it's not just a question
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
where_x, float where_y, float where_z);⏎void deallocate_model(model_3d_ref kill_this);⏎The compiler doesn't read your code like this -
-
Lab found healthy certified buildings pull in effective rents that are 4.4% to 7.7% more per square foot than non-certified and non-registered peer buildings.⏎What’s good for the planet and what’s good for people
-
The reason? Google and its ever-changing algorithms. Specifically, one of the latest actions Google had taken to fight fraudulent, insecure websites and pages is to label non-HTTPS pages as “not secure”:⏎There’s a high chance you might have
-
that are IC rated and those that are not IC rated. The difference between the two ratings is that the IC rated fixtures can be buried in insulation and the non-IC rated fixtures cannot be covered or in contact with insulation. The reason for this
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
- L4:53 excites blank slot before a number
- L0:3025 weakly excites
- L4:1056 weakly excites
- L3:820 weakly excites
- L3:2152 weakly suppresses
- L2:3028 weakly excites
- L3:2759 weakly suppresses
- L4:859 weakly excites
- L4:71 weakly suppresses
- L4:2838 weakly excites
- L3:2009 weakly suppresses
- L3:858 weakly excites
- L2:944 weakly suppresses
- L0:2391 weakly excites
- L4:232 weakly suppresses
- L0:231 weakly suppresses
- L2:2192 weakly suppresses
- L4:2344 weakly suppresses
- L3:1018 weakly excites
- L2:543 weakly excites
- L2:424 weakly excites
- L4:1748 weakly suppresses
- L3:2111 weakly suppresses
- L2:716 weakly suppresses
Feeds later layers
- L14:140 excites
- L11:300 weakly suppresses
- L7:8 weakly suppresses
- L16:396 weakly excites
- L19:1351 weakly suppresses
- L19:786 weakly excites
- L18:2578 weakly excites
- L8:162 weakly suppresses
- L8:111 weakly excites the space before a year in dates
- L19:1254 weakly excites
- L13:2166 weakly suppresses
- L17:1161 weakly suppresses
- L12:613 weakly excites
- L24:197 weakly excites
- L6:152 weakly suppresses
- L26:1587 weakly excites
- L18:76 weakly excites blank spot before a number
- L26:840 weakly suppresses
- L20:59 weakly excites
- L22:164 weakly suppresses
- L6:2828 weakly suppresses
- L6:2 weakly suppresses
- L22:52 weakly excites
- L14:348 weakly 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
- ColumnType
- tures
- ortal
- geber
- agnar
- cctor
- onent
- FTER
- вещ
- 治理体系
Suppresses
- ec
- fe
- etest
- ้น
- ử
- Enlarge
- eload
- masks
- ep
- suppress