digits inside citation and DOI strings
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. 22 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 the numbers and punctuation inside reference-style strings, like the digits in a journal DOI or a page range in a bibliography. The steering results did not show a clear theme. Turning it up made the text repeat words and phrases, and turning it down gave repeated number patterns.
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.
-
, 155. 50-58. 10.1016/j.biocon.2012.05.013Full text not available from this repository.⏎Although grasslands are crucial
-
-1600. http://dx.doi.org/10.1016/j.jbusres.2013.10.001⏎BERISHA-NAMANI, M. (20
-
.⏎"Purchasing Power Parity Exchange Rates for the Global Poor,"⏎American Economic Journal: Applied Economics,⏎American Economic Association, vol. 3(2), pages 137-66, April.⏎- Angus Deaton & Olivier
-
13, 5(7), 3006-3008; doi:10.3390/su5073006⏎I am honored to introduce this special issue of Sustainability,
-
with Nupecoupé, toupée⏎What do you find interesting about this word or phrase?⏎Comments that don't adhere to our Community Guidelines may be moderated or removed.<|endoftext|>Why travel to Europe to see old ruins when you have some
-
Medicine study are available here: http://www.europeangeriaticmedicine.com/article/S1878-7649(13)00123-X/).⏎It is possible nursing staff underestimate the importance of oral
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: in the context of the real-tn-2011-12-24-12-25-12-26-12-27-12-28-12-2
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
; CIA, 2003). Furthermore, parks and<|endoftext|>From: Mary Ann Liebert Inc.⏎Posted: Saturday, September 1, 2007⏎Astrobiology August 2007, 7(
-
sickly, thin, or emaciated appearance; drawn; said of the face or the expression.⏎from WordNet 3.0 Copyright 2006 by Princeton University. All rights reserved.⏎- adj. somewhat ill or prone
-
from Wiktionary, Creative Commons Attribution/Share-Alike License⏎- n. An underwater explorer.⏎from WordNet 3.0 Copyright 2006 by Princeton University. All rights reserved.⏎- n. an underwater swimmer
-
two different numbers cant have the same cube). Therefore, by Definition 1, is one-to-one....⏎View Full Document⏎- Spring '09<|endoftext|>part of speech: noun⏎definition 1: an upright line
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
- L20:790 excites
- L19:1627 excites
- L18:2060 suppresses
- L19:619 suppresses
- L20:1342 excites
- L13:1606 excites
- L19:1017 weakly excites
- L18:764 weakly suppresses
- L11:1290 weakly excites
- L21:362 weakly excites
- L19:1242 weakly excites
- L0:730 weakly suppresses
- L19:279 weakly excites
- L19:2153 weakly suppresses
- L20:1090 weakly excites
- L14:2170 weakly excites
- L20:806 weakly suppresses
- L19:616 weakly excites
- L4:21 weakly excites
- L18:1092 weakly excites
- L21:2799 weakly suppresses
- L20:3046 weakly excites
- L21:2030 weakly excites
- L5:364 weakly suppresses
Feeds later layers
- L23:363 strongly suppresses
- L24:710 suppresses
- L25:2507 suppresses
- L23:349 excites comma after author surname in citations
- L26:386 excites
- L23:404 suppresses
- L23:2370 suppresses
- L26:1756 suppresses
- L27:268 suppresses
- L23:383 suppresses
- L23:1206 weakly excites
- L26:219 weakly suppresses
- L26:539 weakly suppresses
- L25:915 weakly suppresses
- L27:1361 weakly suppresses
- L27:509 weakly suppresses
- L24:2045 weakly suppresses acronym introductions
- L23:1690 weakly suppresses
- L24:749 weakly suppresses
- L27:295 weakly excites
- L25:522 weakly suppresses
- L27:468 weakly excites
- L23:531 weakly suppresses
- L25:1223 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
- ớt
- Tower
- Bellev
- Presidents
- =v
- Mods
- лага
- `%
- 扞
- 很正常
Suppresses
- Ak
- Ci
- 堪
- Ec
- Fn
- 兼
- Ai
- Ai
- tn
- Mn