name fragments inside split 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. 26 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 weakly on a piece of a word that has been chopped up, mostly inside proper names, such as "ust" in "Eustis", "igr" in "Tigris" or "Car" in "Carpathia". The examples are varied and the signal is weak, so the pattern is not very sharp. Steering changes did not reveal a 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.
-
every titan finds its challenger eventually. This one just happened to appear in the form of a bee swarm.⏎The drama unfolded Thursday at the Joint Base Langley-Eustis in Virginia, when an F-22 aircraft from the 19
-
of different kingdoms spread out through the land between Anatolia (modern Turkey) and Babylon (modern Iraq).⏎Mitanni took brief control of the region in the Tigris and Euphrates basin between the fall of Amorite Babylon to
-
formed in 1880 from part of the Town of Peirrepont, was the last town formed in the county. Madam De Stael, a French author, at one time invested in the town lands, on
-
before you can see the other students’ posts. Submit your initial post(s) early in the module.⏎See the Discussion Board Guidelines and Rubrics⏎Download Discussion Board Guidelines and Rubrics.⏎Reh, F. J. (201
-
journalistic institution, Adafruit strives to be a fair, informative, and positive voice within the community – check it out here: adafruit.com/editorialstandards⏎Stop breadboarding and soldering – start making immediately! Adafruit
-
This is one of Atlanta’s roughest areas.⏎This mansion in the Vine neighborhood, built in 1910, was the home of Alonzo Herndon, who was born into slavery but went on to become Atlanta's first black
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: option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option option
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
earlier version, you should certainly upgrade; this version fixes quite a few problems, including handling of certain kinds of global variables.⏎The first step in creating a GLib shared library is to create a ``static'' library (also called an ar archive
-
Acceptable is a client error that is returned by the server to indicate that the resource exists but the proactive Content Negotiation engaged in by the client using a relevant HTTP request header was not satisfiable and the server does not supply
-
treatment. Prognostic and predictive factors are often discussed together. They both play a part in deciding on a treatment plan and a prognosis.⏎The following are prognostic factors for bile duct cancer
-
show that arable reversion is an effective tool for the management of butterflycommunities. We suggest that as restoration takes time to achieve, its use as a mitigation tool against future environmental change (i.e. by decreasing isolation in fragmented landscapes)
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
- L20:2117 excites
- L22:739 suppresses 'are' after 'here' introducing a list
- L21:913 suppresses
- L19:2379 suppresses
- L25:399 suppresses
- L20:928 excites
- L22:186 excites
- L25:320 excites
- L25:374 suppresses
- L18:872 weakly excites the digit 8 in lists and numbers
- L12:1507 weakly excites
- L17:1172 weakly suppresses
- L22:426 weakly excites
- L4:2743 weakly excites
- L19:2149 weakly excites
- L18:2147 weakly suppresses
- L3:1331 weakly suppresses
- L19:2305 weakly suppresses
- L10:641 weakly suppresses
- L25:437 weakly excites
- L19:2122 weakly suppresses
- L24:451 weakly suppresses
- L25:390 weakly suppresses
- L6:821 weakly excites
Feeds later layers
- L27:551 suppresses
- L27:814 suppresses
- L27:588 suppresses
- L27:1035 excites
- L27:488 weakly suppresses
- L27:1181 weakly excites
- L27:480 weakly excites
- L27:592 weakly suppresses
- L27:1218 weakly excites
- L27:533 weakly suppresses
- L27:2939 weakly suppresses
- L27:584 weakly suppresses
- L27:250 weakly suppresses
- L27:602 weakly suppresses
- L27:1313 weakly suppresses
- L27:403 weakly suppresses
- L27:2567 weakly suppresses
- L27:963 weakly suppresses
- L27:2909 weakly excites
- L27:2806 weakly suppresses
- L27:1927 weakly suppresses
- L27:2387 weakly excites lowercase m opening a word fragment
- L27:1405 weakly excites
- L27:1186 weakly excites
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
- ections
- owi
- Methods
- ddb
- 各项工作
- Methods
- ewan
- methods
- 毯
- behaviors
Suppresses
- option
- (option
- _option
- requirement
- option
- property
- Option
- limitation
- principle
- capability