digits starting a percentage or rate
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. 12 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 fires on the first digit of a number that is about to be a percentage or rate, such as the 5 in '5%' or the 6 in '6.7 percent'. It is strongest when the number is part of a rate or proportion. Turning it up pushed the writing toward numbers and rates, but the text was often confused.
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.
-
follow these instructions.⏎The oxygen content should be more than 5% in any condition. Aerobic composting needs a large amount of oxygen during
-
to receive a positive result than were persons 10 years of age or older, with percentages of 6.7 percent and 13.7 percent, respectively, for targeted testing
-
11:37 am by Matt Prescott<|endoftext|>A significant fraction (~ 30%) of the gamma-ray sources detected by the Fermi Gamma-ray Space Telescope
-
981. Used by permission of the National Council of Teachers of Mathematics.⏎merical values and use it when x=3 (with success rates of 69% and 81% for the two groups of students, respectively).
-
uses of this calculation is when purchasing a product on sale.⏎Thus, if the marked price is 1,000 and has a discount rate of 25% should take the following relationship: 1.000×2
-
in dizygotic twin pairs. Multifactorial threshold models suggested that the heritability of liability to alcoholism in women is in the range of 50
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 answer to a question that I have not been thinking about. The answer to the question that I have not have been thinking about. The answer to the question that I have have not have been thinking about. The answer to the question that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Culture and Sports as part of the commemoration of the 2500th anniversary of the Battles of Thermopylae and Salamis, offers
-
the monument. Where it is not available, the satellite view is shown instead.⏎Latitude: 52.5162 / 52°30'58"N⏎Longitude: -3.0604 /
-
5 alone, AIDS killed 1.1 million people, 800,000 of whom were living in sub-Saharan Africa. Though life-saving antiretrov
-
possible answer to this question in a study published in the journal Science. Based on their analysis of duplicate gene<|endoftext|>The maternal mortality rate reported in Tibet (400/100,000) is higher than that found in most
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
- L10:672 suppresses
- L11:2193 suppresses
- L10:1234 suppresses
- L11:2377 suppresses
- L4:1450 excites
- L11:2352 excites
- L9:2003 suppresses
- L10:635 weakly suppresses
- L5:1764 weakly suppresses
- L9:1222 weakly excites
- L10:1525 weakly excites
- L1:1464 weakly excites
- L10:2268 weakly suppresses
- L4:544 weakly excites
- L9:1723 weakly suppresses
- L9:2526 weakly excites
- L10:2300 weakly excites
- L8:2685 weakly suppresses
- L10:2709 weakly excites words that introduce a clause
- L6:646 weakly excites
- L9:2650 weakly suppresses
- L11:2967 weakly excites
- L9:1884 weakly excites
- L9:2915 weakly excites
Feeds later layers
- L16:1466 suppresses
- L22:2003 suppresses
- L16:2551 suppresses
- L13:2296 excites
- L22:1494 suppresses
- L22:1638 suppresses
- L23:1079 weakly excites
- L24:992 weakly excites
- L17:2230 weakly excites
- L22:2815 weakly suppresses
- L16:2270 weakly excites
- L16:2478 weakly excites
- L19:422 weakly suppresses
- L14:2257 weakly excites
- L20:238 weakly excites
- L14:542 weakly excites
- L17:161 weakly excites blank spot where a number goes
- L13:1397 weakly suppresses
- L18:1114 weakly excites
- L24:2097 weakly excites
- L21:1749 weakly excites
- L21:1103 weakly excites
- L16:1060 weakly suppresses
- L22:1552 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
- sure
- iceps
- 百分
- 刎
- %C
- _CBC
- %A
- anzeigen
- addCriterion
- _syntax
Suppresses
- brero
- ʊ
- ��
- IAS
- arel
- .sl
- 科教
- waiting
- reaching
- 溷