comma after qualifying adverbs
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. 3 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 fires on the comma that comes right after a short phrase like "if so,", "or worse," or "more commonly," where the writer is about to add a qualifier or twist. When it was turned up, the text wandered off into unrelated rambling rather than showing a clear 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.
-
church traditions where Lent is not observed at all. These folks may think of Lenten observance as, at best, a hollow custom, or, at worst, quite foreign to authentic Christianity. As a matter of fact, many who grew up
-
of conduct on those who dissent from them; to fetter the development, and, if possible, prevent the formation, of any individuality not in harmony with its ways, and
-
ate winter look!<|endoftext|>While doing my own comparison of women in the New Testament and Josephus’ Jewish histories, I began to notice a theme, or rather, two themes. Each one dealing with customs involving women and since I could not find
-
tolls than do wealthier people. But in the US, road systems are financed primarily through fuel taxes, vehicle registration fees, property taxes, and, increasingly, sales taxes—all of which are also regressive. Thus the relevant question
-
in Scotland originated with Anglo-Norman knights bearing the de Beauville or, more commonly, de Boyville name from Beauville, near Caen. They are believed to
-
any of these, you may have dry eye. A thorough ocular surface evaluation with our dry eye specialists will determine whether dry eye is present, and if so, what type(s) of dry eye are affecting
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 am" by Emily Brontstern? A. E. Brontstern A. E. Brontstern A. E. Brontstern A. E. Brontstern A
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Men are 24 % less likely than women to visit their doctor for an annual checkup, even though their cancer mortality rates are far higher.⏎Moreover and as a matter of fact about men, let’s be honest. No man looks
-
fitness, it has also been shown to have other health benefits, including mental health benefits such as helping to manage depression and anxiety. Most importantly for many people, running is a good way to lose weight and generally get fitter
-
straight edge. the scrap ends left over from cutting a square from a rectangle work perfect. they can be used in two different ways, fractionally, and additively.⏎when working on any piece, it’s helpful to know where
-
une after bloom in late spring or summer. Wait until all chance of frost is past before pruning tropicals such as bougainvillea and hibiscus. Pruning stimulates tender new growth, which is susceptible to frost damage.⏎
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
- L0:2633 weakly suppresses
- L2:86 weakly suppresses
- L1:198 weakly suppresses
- L2:2940 weakly excites
- L1:2336 weakly excites
- L2:1809 weakly suppresses
- L0:2526 weakly suppresses
- L0:2495 weakly suppresses
- L0:224 weakly suppresses
- L1:2353 weakly suppresses
- L1:2417 weakly suppresses
- L1:465 weakly excites
- L0:1873 weakly excites
- L1:1426 weakly suppresses
- L0:3013 weakly suppresses
- L2:3062 weakly excites
- L2:2425 weakly excites
- L0:1342 weakly excites
- L2:2915 weakly excites
- L2:670 weakly excites
- L1:1889 weakly suppresses
- L2:1036 weakly excites
- L1:2625 weakly excites
- L0:1189 weakly suppresses
Feeds later layers
- L4:1444 excites
- L6:918 weakly excites
- L8:611 weakly excites
- L15:2835 weakly suppresses
- L9:2771 weakly suppresses
- L5:540 weakly excites
- L6:2575 weakly excites
- L4:1441 weakly excites
- L4:2610 weakly suppresses
- L12:2332 weakly suppresses
- L17:794 weakly suppresses
- L13:1706 weakly suppresses
- L20:2065 weakly suppresses
- L17:2891 weakly excites
- L16:1993 weakly suppresses
- L15:396 weakly excites
- L6:2147 weakly suppresses
- L9:2769 weakly suppresses
- L8:852 weakly suppresses comma inside large numbers
- L21:1643 weakly excites
- L4:610 weakly suppresses
- L6:137 weakly suppresses
- L4:744 weakly suppresses
- L4:810 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
- essenger
- semblies
- plementation
- ístico
- ово
- úmero
- asta
- PLICATION
- structor
- plement
Suppresses
- etine
- 蓐
- ()`
- /$',
- ()'
- %'
- ()%
- ?>"><?
- `⏎⏎
- apanese