Checkout my ChatXML proposal. I recommend trying out the prompt generating prompt example. It works best with GPT-4 but still works ok with GPT-3.5. Assistant Definitions as defined all work well with every language model I've tried.
#ChatXML - Highly Structured Prompting
deep cypress
It's also my theory that OpenAI is headed down a similar path with their ChatML approach.
If I am right, then ChatXML should only grow more powerful as they RLHF their way to a structured mark down language.
ebon steppe
Hey friend I'm actually trying to document these novel prompt engineering techniques as they arise. Would you mind if I whipped up a quick blog about this? I can DM it to you so you can provide feedback before I post
ashen phoenix
DM me a link to your blog site I would like to take a read
ebon steppe
so the only one I have up write now is my approach...I have one drafted for Quicksilver OS but waiting on Quicksilver to give it approval
if you go to blog_synapticslabs_ai you can check them out (replace the underscores with periods)
peak hedge
looks like an interesting project to study, I'm already reading some studies and articles about it, if you have any specific content that talks about please share it
humble totem
Checkout my ChatXML proposal. I recommend trying out the prompt generating promp...
Have you noticed any benefits to using XML structure instead of JSON? From a tokens standpoint JSON is a better choice because it doesn't require the closing tag for each element.
peak hedge
Have you noticed any benefits to using XML structure instead of JSON? From a tok...
I've tested about 3 usage implementations with Json and Chatxml. But what I've noticed so far is that while Json can be broader, while chatxml is more focused, clearer and more readable than json, but you have to see if chatxml is more focused, specifically defining what should be done, don't end up causing errors by being restricted to act and respond in only one way. (What I've noticed so far, I'm looking at the creator's github and looking to see if I can find other comments about it)
deep cypress
Have you noticed any benefits to using XML structure instead of JSON? From a tok...
Yes I started with a json-like aproach and switch to XML. I think the tag names are the deciding factor. The tags themselves help prompt. It's also a lot better at parsing nested information using the XML. Json seemed to leak since it only had the curly brackets to figure out the nest from.
The delta in tokens is honesly pretty small, I don't mind paying for it.
deep cypress
Hey friend I'm actually trying to document these novel prompt engineering techni...
Go for it! The more structured prompting is discussed and tested, the better imo. I really want to see the community move away from adjective spaghetti and towards something more rigorous.
deep cypress
Have you noticed any benefits to using XML structure instead of JSON? From a tok...
This was an entirely qualitative decision on my part. One day I'll actually test it out.
In fact if you check out the rest of my github you'll see that the assistant definitions are still stored in json instead of markdown because that's what I started with and I am more comfortable using it.
But from my qualitative perspective is that the tags are the way to go.
humble totem
This was an entirely qualitative decision on my part. One day I'll actually test...
Interesting. Thanks for the insight.
ebon steppe
Go for it! The more structured prompting is discussed and tested, the better imo...
Just sent a friend request I have a google doc drafted for you to check out
ebon steppe
Hey all blog is posted. Can’t hyperlink in here but just go to blog _ Synapticlabs _ ai (_=.)
deep cypress
Hey all, I made some updates to the repo. Added a discussion tab, readME and most importantly a prompts list. Feel free to share any ChatXML prompts that yall develop and I will feature them on the repo.
It's on my github - bradybry.