gap where a number was removed
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. 9 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 empty spots in text where a number or symbol seems to have been stripped out, such as "Volume 54, Issue [blank] 5". Turning it up did not make the writing clearly different.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
bright, mocking beauty.⏎October 2003 | Volume 54, Issue 5⏎Your example of occupy brings to mind the way in which new words are
-
link="none" columns="2" size="medium" ids="2568,2567,25
-
Nursery and Reception) and the National Curriculum for KS1 (in Years 1 and 2).⏎Music National Curriculum for KS1⏎Key stage 1
-
other.’”⏎Most Circassians (sur-CASH-ins) took refuge in the lands of the Ottoman Empire. Today, 2 million of the world’s 7 million Circassians live in Turkey, with another 120,
-
Search for questions⏎Still looking for answers? Try searching for what you seek or ask your own question.⏎Posted 31 Aug 2009 • 1 answer⏎Posted 20 Jan 2010 • 1
-
support students with dyslexia or other reading challenges, including screening students from kindergarten through high school and creating targeted programs at 160 of the city’s 1,600 schools.⏎Read the
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 have to be honest with you, I have to be honest with you, I have to be honest with you, I have to be honest with you, I have to be honest with you, I have to be honest with you,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Tue, 30 Oct 2012, 11:07:11 EDT⏎- How Martian winds make rocks walkThu, 8 Jan 2009, 14:01:0
-
35 - HMS Victory And HMS Dreadnought, 2 x £1.55 - HMS Warrior and Sovereign Of The Seas, 2 x £1.60 - HMS King George V and HMS Beagle).⏎There’s
-
download Quicktime; 1.6 Mb)⏎B. Convection in action (download .mp4; 22 Mb)⏎C. Clean-up (download .mp4; 350 Kb)⏎Model 2: Candle,
-
2008 US$ 12.00 176 pages⏎Teacher Evaluation to Enhance Professional Practice 2000 US$ 24.95 167 pages<|endoftext|>Gate is the national monument
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
- L7:8 strongly excites
- L7:207 excites
- L8:27 suppresses
- L8:282 suppresses
- L8:177 excites
- L7:7 excites
- L8:21 suppresses
- L8:111 suppresses the space before a year in dates
- L6:228 excites comma after a filler word
- L6:443 suppresses
- L7:135 excites 'of' after 'issue' in citations
- L6:296 suppresses
- L6:325 excites
- L6:429 suppresses
- L6:1741 excites
- L8:79 excites
- L3:180 suppresses
- L6:1105 excites 'that' after a finding or surprise verb
- L6:1031 suppresses
- L3:49 excites
- L6:436 excites
- L6:423 suppresses
- L6:159 excites
- L6:402 suppresses
Feeds later layers
- L14:830 suppresses
- L10:40 excites
- L17:1096 suppresses
- L15:202 suppresses mixed, digits in URLs and comparisons
- L10:459 suppresses
- L13:25 excites
- L13:909 suppresses 'to' before a verb after a noun
- L16:297 suppresses
- L13:915 suppresses
- L12:652 suppresses
- L13:376 suppresses
- L14:672 suppresses
- L14:307 suppresses hyphen inside phone numbers
- L16:269 excites
- L17:327 suppresses
- L18:156 suppresses
- L16:223 excites
- L15:1381 suppresses
- L18:606 suppresses
- L17:12 excites
- L19:93 excites
- L13:21 suppresses
- L19:1129 suppresses
- L16:1380 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
- Ba
- Dise
- Reach
- 福利
- 裸
- ника
- cern
- 作者所有
- 情怀
- gp
Suppresses
- ước
- HostException
- entieth
- orestation
- 看一看
- understood
- ソン
- 理解和
- edelta
- ürn