ordinal numbers in military unit names
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. 9 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 the "th" in numbered military units, like the 39th Fighter Squadron or the 84th Indiana Infantry. When it was turned up, the model's writing drifted toward soldiers, armies and the military.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
-86 Sabres took place in the skies over North Korea. Captain Joseph C. McConnell, Jr., leading a flight of fighters of the 39th Fighter Interceptor Squadron, 51st Fighter Interceptor Wing, based at
-
began to recruit a second all-black regiment of soldiers to join the fight.⏎As the struggle against the Confederate South raged on, the newly-formed 54th Massachusetts battalion prepared to attack the seemingly impregnable fortress of Fort Wagner. The
-
Army. Indiana Infantry Regiment, 84th (1862-1865)⏎Historical Note: The
-
May 14, 1860.⏎With the coming of the Civil War, Tinkham helped to organize Company F of the 26th Illinois Volunteers regiment and was commissioned as Captain on August 10, 1
-
allergic reactions, you should talk with your doctor and request a vaccine with the least egg protein available<|endoftext|>United States. Army. Indiana Infantry Regiment, 84th (1862-1865) | Purdue University Libraries,
-
of New Alexandria to learn the blacksmith trade before enlisting in the Union Army.⏎In 1861, William enlisted in the Pennsylvania 11th Reserve Infantry. He was just 18 and small in stature. Military officials
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: when I look at the graph of a function, say, a line, I can see that the function is a curve that crosses the x-axis. If I use a calculator to find the value of the product of two numbers, say,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. An experimental 500-w laser built for the U.S. Army was field-tested successfully, destroying 98% of land mines and unexploded mortar shells in only a few minutes. The laser ignited the explosives by heating their
-
costs about $148 million.⏎Needless to say, the F-22 is kind of a big deal. Going head to head with the multi-role fighter under any circumstance sounds like a bad idea, given that its primary function is
-
Users, 2012-2017). According to Foreign Policy, The Defence Advanced Research Projects Agency is developing some extremely high technology to tackle possible alien invasions, an example being
-
ments, particularly following Hitler's infamous ‘Commando Order’ issued on 18 October 1942, ordering the annihilation of all men operating on Commando raids against
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
- L8:2626 excites
- L6:2995 excites
- L7:2377 weakly suppresses
- L5:133 weakly excites
- L6:873 weakly excites
- L6:2332 weakly excites
- L8:10 weakly excites
- L3:2294 weakly excites
- L4:2923 weakly suppresses
- L2:554 weakly suppresses
- L5:2085 weakly suppresses
- L5:2135 weakly excites
- L6:1608 weakly excites
- L7:1513 weakly excites
- L7:2379 weakly suppresses
- L7:2366 weakly suppresses
- L8:89 weakly excites
- L4:1700 weakly excites conservation laws and invariance
- L5:495 weakly suppresses
- L2:3070 weakly excites
- L5:57 weakly suppresses
- L3:916 weakly excites
- L6:1459 weakly excites
- L6:1774 weakly excites
Feeds later layers
- L18:2097 excites
- L20:2387 excites
- L24:2720 excites
- L19:2937 excites unclear: mixed possessives and fragments
- L16:1460 excites 'of' after a group noun
- L11:1519 excites
- L19:2473 excites
- L12:1348 suppresses
- L22:1196 excites
- L17:978 excites
- L21:372 weakly excites
- L10:2608 weakly excites
- L10:2261 weakly suppresses
- L16:2518 weakly excites
- L21:3001 weakly excites
- L10:1880 weakly excites
- L21:1831 weakly excites
- L11:1941 weakly excites
- L19:2550 weakly suppresses
- L23:1669 weakly suppresses
- L17:2150 weakly suppresses
- L24:2772 weakly suppresses
- L20:1247 weakly excites
- L18:2997 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
- 队伍
- 队
- 军队
- Battalion
- 班组
- crew
- Milit
- 部队
- illery
- military
Suppresses
- egg
- emark
- nues
- GetProperty
- emodel
- bart
- [item
- egot
- ULATOR
- 妣