page and range numbers in web text
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. 10 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 numbers that point to pages, ranges or zones in web boilerplate, such as 'Pages 4 (1004 words)' or 'pages 1-2'. It also fires on some numbers in citations. Steering changed the topic in every direction, so it doesn't show a clear effect on what the model writes.
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.
-
: AL-lee-um ka-ra-tah-vee-EN-see⏎Sunset zones 1-24.⏎USDA zones: 5-9.⏎Heat zones: 9-5.⏎Height: 4-
-
caused<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally blurred
-
this week, the jury is still out on that one.<|endoftext|>Watch: 1929 Stock Market Crash and the Great Depression – Documentary [58:35 minutes]⏎After watching the above video and the relevant discussion in Special
-
How to Read Music Notes - Introduction⏎No comments yet.<|endoftext|>Please boost your Plan to download papers⏎Essay example - The music of South India⏎Pages 2 (502 words)⏎This essay describes what are the basics of Indian music
-
garbage patch, also described as the Pacific trash vortex, is a gyre of marine debris particles in the central North Pacific Ocean located roughly between 135°W to 1
-
water to further alleviate local flooding.⏎Throughout its preparation for and response to Isaac,<|endoftext|>Please boost your Plan to download papers⏎Exercise Behavior in Adults⏎Pages 4 (1004 words)⏎Although the benefits of regular physical activity have been
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 can't do it right. I have to do it in a way that's important. It is not possible to use tools, and I can't. I have to use a tool. It is not possible to use a tool,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Anthony and Minneapolis. Other large homes of merchant families within the Fifth Street Southeast area include the Van Cleve residence (603 5 th St. SE), William McNair (610 6 th St. SE) and
-
vibrant small community founded in 1825 by free working class African-Americans in uptown Manhattan. The area from W82nd to 88th Streets between Seventh and Eighth Avenues (click here for a larger version of the
-
yliving/antioxidants⏎- Vitamin C and Immune Function, 2017, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5707683/#:~:text=It%2
-
iodiversity and Conservation (2004) 13: 2511. doi:10.1023/B:BIOC.0000048451.07820
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
- L7:1196 suppresses
- L8:729 suppresses
- L8:785 suppresses
- L9:1224 excites
- L6:411 suppresses
- L9:515 excites
- L5:2823 excites
- L5:966 weakly suppresses
- L9:730 weakly suppresses
- L9:1243 weakly excites
- L2:398 weakly suppresses
- L5:1136 weakly excites
- L5:884 weakly suppresses
- L4:1435 weakly excites
- L8:2744 weakly excites
- L7:2342 weakly suppresses
- L1:2629 weakly suppresses
- L3:327 weakly excites
- L6:2076 weakly excites
- L5:1373 weakly suppresses
- L7:2337 weakly suppresses
- L0:182 weakly suppresses
- L5:1605 weakly excites
- L6:908 weakly suppresses
Feeds later layers
- L11:1042 excites
- L16:2776 excites
- L11:703 excites
- L11:598 suppresses
- L16:2121 suppresses
- L11:885 suppresses
- L11:1160 excites
- L13:996 excites
- L11:912 excites
- L16:1676 excites
- L11:1560 excites
- L13:1090 suppresses
- L12:487 excites
- L11:805 suppresses
- L18:1699 excites
- L13:1183 suppresses
- L16:1536 excites
- L16:2519 suppresses
- L13:1895 suppresses
- L11:3071 excites
- L21:1628 suppresses
- L13:2737 weakly suppresses
- L11:954 weakly suppresses
- L12:815 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
- arding
- ced
- ending
- 位
- ese
- orne
- Parsing
- ija
- 心想
- postponed
Suppresses
- ulan
- aneous
- ycz
- cimal
- ูล
- acidad
- ką
- irical
- Mistress
- datal