hyphens inside codes and page ranges
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 dash in things like standard numbers, page ranges, and gene or product codes, for example 1877-1882 or miR-210. When it was turned up, the replies did not show a clear shared theme.
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.
-
are and point out which one will work better⏎Why is ventilation used in buildings?⏎According to the Polish standard PN-83/B-03430, every building must be equipped with
-
flow rate of 16 l/min was used for shielding throughout the welding.⏎Some of fabriors have had success welding galvanized tube using E70C-6 metal cored wire such as Hobart’s Galvacor; the parameters
-
a. The cluster is called PSZ1 G311.65−18.48, and the lensed galaxy is PSZ1- 36 ARC G311.6602–18
-
way to elucidate the molecular basis of the genetic susceptibility to alcoholism.(JAMA. 1992;268:1877-1882)<|endoftext|>The Imaginative Treatment of Geometry in Waldorf
-
: Saturday, September 1, 2007⏎Astrobiology August 2007, 7(4): 578-604⏎The European Space Agency's ExoMars mission will seek
-
pushing in all directions. It is also "jetting" out from the mouth of the balloon. The end of the<|endoftext|>Crystallography, cryo-EM and AlphaFold: a perfect trio sheds light on key human antibacterial proteins
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 old way, and I am to be a person. I need to hear the old way. I am to be a person. I need to hear the old way. I am to be a person. I need to hear the old
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Zelmanov spoke up first, saying, “A proof is what is considered to be a proof by all mathematicians, so I’m pessimistic about machine-generated proofs.” He mentioned the four-color theorem
-
, as illustrated in the picture.⏎What's more, the model shows that a strange phenomenon might occur in the uppermost liquid layer, where ice floats upward. In this scenario, cold plumes cause Ice III to form. As
-
uble vitamin -- others are A, E and K -- vitamin D can be stored in the tissues of your body. This is in sharp contrast to water-soluble vitamins, excess quantities of which are flushed from the body as waste products. Monitor
-
in multiple layers, more like a club sandwich.⏎Ice comes in different forms depending on pressures. "Ice I," the least dense form of ice, is what floats in your chilled beverages. As pressures increase
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
- L4:799 suppresses
- L4:46 suppresses
- L3:101 suppresses
- L4:1056 weakly excites
- L4:303 weakly suppresses
- L3:573 weakly suppresses
- L3:2106 weakly excites
- L3:2088 weakly suppresses
- L1:198 weakly suppresses
- L3:2993 weakly suppresses
- L3:807 weakly suppresses comma inside dates and years
- L1:2324 weakly suppresses
- L2:1036 weakly excites
- L3:338 weakly suppresses
- L4:1756 weakly excites
- L3:1115 weakly excites
- L2:525 weakly suppresses
- L4:556 weakly suppresses
- L3:2368 weakly suppresses
- L4:2658 weakly excites
- L2:359 weakly excites
- L1:1652 weakly excites
- L2:1824 weakly suppresses
- L3:1821 weakly suppresses
Feeds later layers
- L19:804 excites
- L7:89 excites dash in year or number ranges
- L23:1479 excites
- L17:364 weakly excites hyphen joining two words
- L22:2921 weakly excites
- L17:2269 weakly suppresses
- L7:434 weakly excites
- L24:2124 weakly suppresses
- L17:2348 weakly suppresses
- L24:1957 weakly excites capital 'In' starting a word or name
- L6:1111 weakly excites
- L27:2543 weakly suppresses
- L25:1288 weakly suppresses
- L9:1418 weakly excites
- L19:1547 weakly suppresses
- L21:942 weakly suppresses
- L10:2686 weakly excites
- L20:449 weakly suppresses
- L6:1055 weakly excites
- L6:2812 weakly suppresses
- L22:2026 weakly excites
- L9:2932 weakly excites
- L17:206 weakly excites
- L13:2378 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
- Capability
- elig
- гин
- 信誉
- 私立
- �
- atics
- yna
- 跛
- creep
Suppresses
- зер
- ave
- FSIZE
- marshal
- rometer
- ughters
- ilar
- scaled
- assel
- kowski