'of' after quantity words
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. 2 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 word 'of' when it follows a word about an amount, such as 'a large percentage of', 'the majority of' or 'a fraction of'. Turning it up did not make the model's text clearly reflect this pattern.
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.
-
the nutrition labels on the products they buy, and a large percentage of them don’t read the labels at all. The question is why?⏎Researchers at
-
vision nor mission statements include the words “aeronautics” or “space”. When the Space Shuttle program ended, a substantial fraction of the general public (along with many news reporters) surmised that NASA was
-
process for creating it.⏎Assessing data health⏎To understand the health of your organization’s data, start by asking questions, such as:⏎- What percentage of data is considered complete?⏎- What percentage of data contains inaccurate, incomplete, or
-
of play-based pedagogies. In addition, a large majority of teaching staff is female.⏎Similar to what research has found to be true in countries
-
incentive mechanism for OEMs to produce zero-emission and low-emission vehicles, involving a relaxation in the specific CO2 emissions of an OEM provided that its share of such vehicles exceeds 15% of
-
alcohol consumption, and poor diet contribute heavily to extrinsic skin aging. But, in most individuals, it is over-exposure to sunlight that accounts for the majority of pre-mature skin aging – up to 90%!⏎Over exposure to
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 a little bit. I have a little bit. I have a little bit. I have a little bit." Let's see, I have a little bit. I have a little bit. I have a little bit. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
) законодательная власть⏎4) Associate Justice d) беспартийная система⏎5) a right of voting e) гражданин государ
-
between exposure to hazardous chemicals and preterm birth in northern Puerto Rico. Particular attention is given to chlorinated volatile organic compounds and phthalates, although biom
-
thritis:⏎Collagen (type II, bovine) is used to induce arthritis in animal studies. These studies are preformed to evaluate anti-inflammatory treatments. Particular mice (a strain called DBA/1) which are particularily sensitive
-
I react negatively to dilation drops, that won’t stop me from having my eyes dilated each visit. I want to prevent any damage to my eyes from cataracts or
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
- L0:474 weakly excites
- L1:1975 weakly suppresses
- L0:1657 weakly suppresses
- L0:1020 weakly excites
- L0:1318 weakly suppresses
- L1:2467 weakly suppresses
- L0:1633 weakly excites
- L1:2151 weakly excites
- L1:62 weakly excites
- L1:1739 weakly suppresses
- L1:1032 weakly suppresses
- L0:2134 weakly suppresses
- L1:2550 weakly suppresses
- L0:2962 weakly excites
- L1:18 weakly excites
- L1:3045 weakly excites
- L0:644 weakly suppresses
- L1:1289 weakly suppresses
- L1:2225 weakly suppresses
- L0:1630 weakly excites
- L0:2147 weakly suppresses
- L0:897 weakly suppresses
- L0:2440 weakly suppresses
- L0:2778 weakly excites
Feeds later layers
- L14:2630 weakly excites
- L5:832 weakly suppresses
- L22:909 weakly suppresses
- L3:2742 weakly suppresses
- L4:1995 weakly excites
- L12:249 weakly excites
- L14:2067 weakly excites
- L4:1361 weakly excites
- L14:597 weakly excites unclear: document ends and list breaks
- L14:26 weakly excites
- L19:550 weakly excites
- L3:709 weakly excites mid-word pieces in anatomy and place names
- L7:1065 weakly suppresses
- L7:2940 weakly suppresses
- L16:1727 weakly suppresses
- L7:1662 weakly suppresses
- L21:1174 weakly suppresses
- L6:1051 weakly excites
- L7:2918 weakly suppresses
- L15:1193 weakly suppresses
- L6:638 weakly excites
- L3:1778 weakly suppresses
- L20:862 weakly excites
- L10:2365 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
- ье
- isers
- doInBackground
- 用户提供
- setError
- endTime
- .getHost
- ьте
- Cosby
- iores
Suppresses
- bu
- alus
- 炮
- token
- fu
- rence
- token
- icht
- sp
- ve