Use Quantitative Metrics to Build Better Prompts.
can be notoriously hard to use. It is common practice to provide these general models context that can push them to act in a way more
The process of aligning large machine learning models to the goals of the programmer is a large and complicated process. Prompts are just one small way to help the alignment process.
We call this context a prompt.The field of prompt engineering is largely based on trial and error. You try things out and see what works. This tool removes the guess work by providing real numbers to predict how well a LLM will behave in a large variety of ways.
These numbers represent how well an LLM will focus on a certain area of knowledge or how patient it will be in dialogue. We are also able to provide custom categories to help you achieve your specific goals.
Disclaimer: Everything below is for illustrative purposes only. There is no live inference, and the categories are not representative of the actual categories in the product.
This doesn't have to be precise, we just use this to figure out what categories may be useful to you.
We have a collection of thousands of curated categories to match your goals,
What you see below is just for illustrative purposes. In the actual demo there are also behavioral categories like "patient" and "aggressive" and the ability to search for any category you want.
Category | Select |
---|---|
Epistemology | |
Metaphysics | |
Logic | |
Ethics | |
Aesthetics | |
Political Philosophy |
We automatically generate a prompt based on the categories you selected, although
We are actively working on improving the prompt generation process, but it is still in its infancy. Expect it get drastically better in the coming months.
Category | Network Recognition | Primary Category |
---|---|---|
Epistemology |
38%
|
Metaphysics |
Ethics |
72%
|
Value Theory |
Aesthetics |
44%
|
Value Theory |
Logic |
38%
|
Metaphysics |
Philosophy of Mind |
83%
|
Metaphysics |
Change the prompt to optimize for the categories and behaviours you are looking for.
Category | Network Recognition | Change | Primary Category |
---|---|---|---|
Epistemology |
50%
|
5% | Metaphysics |
Ethics |
83%
|
14% | Value Theory |
Aesthetics |
33%
|
9% | Value Theory |
Logic |
44%
|
14% | Metaphysics |
Philosophy of Mind |
0%
|
2% | Metaphysics |
We have reverse-engineered parts of a number of to find what parts of the network resonate with different sections of their training
We do not have any special access to the models, we have simply recreated parts of their datasets and recreated subsections of the network. We are confident that this works as they produce similar results to the original models in their given niches.
We are in the proccess of writing an in-depth guide on how everything works and where our current models are lacking, so stay tuned!The basic version is free and will remain free, but we are also working on a premium version that allows greater observation of the model's behaviour and allows users to create their own categories with custom data sets. Beyond that we also offer services to create custom categories for specific use cases as well as creating custom prompts for your specific needs.
We are currentely working with our initial users to improve the general experience. Over the next year, we plan on enabling automatic iteration to generate the best possible prompt for your use case. We are actively listenting to our users and are open to suggestions.