unclear: article ends and dateline openers
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 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 is hard to pin down. Its strongest cases include the end-of-text marker after finished pieces and words like 'article' and 'do' in boilerplate. Its negative cases are the opening bracket after a news dateline. Turning it up gave loose, rambling text with no clear theme.
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.
-
Pakistan⏎||It has been suggested that this article be merged with Blasphemy law#Pakistan. (Discuss) Proposed since
-
/ted/2010/ted_20100805.htm (visited November 27, 2021).<|endoftext|>- Managerial Economics Tutorial⏎- Managerial Economics - Home⏎- Managerial
-
time comtemplating the attraction being caused by curve space when neither were in motion, what causes the force to act upon each other.⏎Hope this is clear enough to get my question across.<|endoftext|>Yesterday I cut myself some fresh pineapple, then promptly
-
tiles has worn, it can't be rejuvenated in any way.⏎Likewise, damage by harsh scratching cannot be reversed.⏎Scratched tiles will continue to perform<|endoftext|>The word biathlon comes from the Greek word meaning two contests. At the Winter
-
marsat, but was withdrawn in 2008 and replaced by the broadband global area network (BGAN), a global satellite Internet network.⏎Techopedia explains Regional Broadband Global Area Network (
-
be their dates. The Sadie Hawkins Dance was popularized by a song of the same name by the punk-rock band Relient K.Learn more about Poetry<|endoftext|>Who has the time or energy to play right now? Our animals do, and
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: that is, in some cases, in the first place, in some cases, in some other cases, in some cases, in some cases, in some cases, in some cases, in some cases, in some cases, in some cases
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Governance⏎- Civil Society⏎Sunday, September 19, 2021⏎MONTEVIDEO, Mar 10 2010 (IPS) - “Nature is wise, and if we take the time to observe
-
akra system (spinning vortexes or wheels of light)? Would you like to know what each chakra represents? Here is a simple diagram and list for
-
retaceous amberJune 16th, 2010 - 1:53 pm ICT by ANI⏎London, June 16 (ANI): Palaeontologists have discovered two mammal hairs encased in amber
-
more than eight top-flight American universities offer degree courses to students from a host of countries.⏎This article cannot deal with every issue on the summit agenda. What follows are a few examples of the challenges and innovations under discussion.⏎Access to Education⏎At
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
- L10:2987 excites
- L8:1984 suppresses
- L11:965 excites
- L10:1910 suppresses
- L9:211 excites
- L11:1740 suppresses
- L7:28 excites
- L10:2284 suppresses
- L10:2821 excites
- L11:2293 excites
- L10:925 suppresses commas in long lists
- L10:1006 suppresses
- L11:633 excites
- L11:1029 excites
- L7:1309 weakly suppresses
- L3:42 weakly excites
- L9:2107 weakly excites
- L6:1460 weakly suppresses
- L2:1804 weakly excites
- L7:732 weakly excites
- L10:300 weakly excites
- L9:591 weakly excites
- L9:252 weakly excites
- L10:1731 weakly suppresses
Feeds later layers
- L14:1411 suppresses
- L13:1817 suppresses
- L14:1727 suppresses
- L16:2724 suppresses
- L15:1003 suppresses
- L16:1010 suppresses
- L18:2489 suppresses
- L13:2498 suppresses
- L22:1900 suppresses
- L13:360 suppresses
- L19:1631 suppresses
- L14:2488 suppresses
- L13:2056 suppresses
- L20:440 suppresses
- L20:388 excites
- L20:2654 suppresses
- L15:2752 suppresses
- L14:1515 suppresses
- L13:3004 suppresses
- L16:2481 suppresses
- L15:1073 suppresses
- L25:882 weakly suppresses
- L17:2093 weakly suppresses
- L14:2753 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
- 涂层
- driv
- 最难
- lung
- forcing
- 增值服务
- กะ
- `),⏎
- etest
- сли
Suppresses
- 为您
- example
- 举例
- examples
- 下面小编
- illustrate
- OffsetTable
- above
- exempl
- opers