weak end of document signal
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 gives a small response at the marker where one document ends and another begins, and a small negative response on single digits inside years. It is weak, so it is hard to say much more. Turning it up gave repetitive or off-topic writing.
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.
-
manage operations in your warehouse.⏎- Lay out a proper strategy for handling damaged products and returns from customers.⏎- Implement pest control measures to protect your products.⏎-<|endoftext|>Nucleation properties of respirable coal dust (terminal progress report).⏎Physics Department
-
will discuss a basic MySQL installation on a Linux virtual server, and some of the basic MySQL commands.⏎Full documentation can be found on the MySQL.com website here.<|endoftext|>The law applicable to military strategic use of outer space⏎New technologies and the law
-
the other hand, are designed to handle complex logic functions that require high performance and low latency.⏎Developing software for microcontrollers is relatively easy compared to FPGA development<|endoftext|>The deepest human need is to engage with the world: confidently, productively,
-
transfer of title by purchase and sale.⏎- It involves mutual benefit both to the buyer and the seller.⏎- It must be directed for earning one’s livelihood.<|endoftext|>Exactly 100 years ago, on 6 January 191
-
00 training examples, you can choose to pick 100 random samples (batch size=100) for each EPOCH (Training cycle)<|endoftext|>|Origin:||co
-
be the most effective approach. Smart work, on the other hand, can save time and increase productivity, but may lead to a lack of focus on the details<|endoftext|>Rajan Vempati of ChK Group, Inc. in Plano,
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: "You can't just ask for a specific question to be answered with the information you need to answer the question properly and provide the user with a clear and well-structured response." "Please follow the instructions for the entire process step by step
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
phoid. Reports of the day state he was frequently asked to describe his techniques before the Ohio Valley Medical Association<|endoftext|>Posted on 09 Aug 2023⏎The arrest of two macaque smugglers in June 202
-
for acyclic graphs of pointers.<|endoftext|>The Death And Life Of The Great Lakes Pdf⏎- and pdf⏎- Tuesday, June 1, 2021 2:11:00 AM⏎- 1 comment⏎
-
5 - November 1, 2020)⏎This introductory course covers the basic history, core principles and challenges of current
-
a potential lunar lava tube entrance – could offer insightful data concerning the moon’s geological history and could hint<|endoftext|>(Updated on June 22, 2020)⏎The mass air flow (MAF) sensor is an important component which
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
- L10:58 strongly excites
- L9:475 strongly excites
- L10:553 suppresses list-style essay phrasing, weak
- L10:80 excites
- L8:642 suppresses
- L9:23 suppresses
- L9:302 excites
- L3:17 suppresses
- L10:1019 suppresses unclear: clause starters in old texts
- L7:1005 excites
- L8:609 suppresses
- L6:2487 suppresses
- L2:604 suppresses the word 'some' before a noun
- L10:1523 suppresses
- L9:1781 suppresses
- L4:1760 excites
- L9:1826 excites
- L4:121 excites
- L8:663 suppresses
- L8:2902 excites
- L3:157 excites
- L9:1544 excites
- L9:614 weakly suppresses
- L8:1462 weakly excites
Feeds later layers
- L13:2100 suppresses
- L17:349 suppresses
- L16:704 weakly suppresses
- L14:71 weakly suppresses
- L16:1957 weakly suppresses
- L26:962 weakly excites
- L27:454 weakly excites
- L25:257 weakly suppresses
- L22:3021 weakly excites
- L23:1613 weakly excites
- L25:2240 weakly excites
- L15:1612 weakly excites
- L17:1587 weakly excites
- L27:1206 weakly excites
- L19:303 weakly suppresses
- L15:1111 weakly suppresses
- L13:626 weakly suppresses
- L17:551 weakly suppresses
- L14:1121 weakly suppresses
- L17:3007 weakly excites
- L27:357 weakly suppresses
- L24:587 weakly excites
- L25:630 weakly excites
- L18:337 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
- ัก
- loth
- ��
- PLIC
- 斋
- uide
- [${
- 化合
- DAYS
- ль
Suppresses
- your
- 您的
- asn
- 你的
- ifton
- anford
- conditional
- */⏎⏎⏎⏎
- >⏎⏎⏎⏎
- innacle