David Shapiro
Patreon: https://www.patreon.com/daveshap (Discord via Patreon)
Substack: https://daveshap.substack.com/ (Free Mailing List)
LinkedIn: https://www.linkedin.com/in/dave-shap-automator/
GitHub: https://github.com/daveshap
Similar Posts
44 thoughts on “Train GPT-3 on Any Corpus of Data with ChatGPT and Knowledge Graphs – SCOTUS Opinions Part 1”
Comments are closed.
Hi David, thank you for starting the tutorial series again. I'd never heard of NLP before stumbling on your channel, and I previously found your tutorials and your way of going through the thinking process incredibly invaluable for a non-coder, non-expert in this domain. I'm excited to see your future projects/videos, keep up the great work.
Thank you for this, you've expanded my understanding greatly.
Amazing stuff ! Your channel is under rated, looking forward the next videos
i’ve learned a lot about spiritually from you which is mad weird for a python channel
Hello, Im the colombian lawyer that you asked for in 10:47 LOL
One of the benefits that I have read of is memory of the past prompts and responses within an instance.
How would you test and perhaps front load context and few facts to improve output?
in France right now and loving your videos!
How has this only got 4k views.
20:00 LOL I lived in France for years and I loved to say that the French and Americans are exactly alike: They both think they are the best(at everything)! 🙂
Love it. I am here for it. Thank you for your leadership. Please continue to lead and inspire us. I wish you the perfect new year. Lisa.
I’ve noticed that specifying a word count improves the specificity. Like if I tell it to give me a response that is at least 2500 words
8:41 actually you can use another AI toy to scrap automatically the whole webpage and sort it as you wish.. it's called "Browse AI"
Great video. Thanks for sharing your thoughts. Learned a lot.
13337 views
why aren't you just going to Google Scholar and clicking 'case law'?
Ok speaking of developer stories, here are some I asked ChatGPT to provide about itself:
1. As a user, I want to be able to provide specific context to ChatGPT so that its responses are more accurate and relevant to my needs.
2. As a developer, I want to be able to train ChatGPT on my own data sets so that it can better understand my company's specific industry and terminology.
3. As a user, I want to be able to easily switch between different versions of ChatGPT, such as a more casual conversational model or a more formal business model, depending on my needs.
4. As a developer, I want to be able to access and analyze the data on how ChatGPT is being used, so that I can improve its performance and make it more user-friendly.
5. As a user, I want to be able to communicate with ChatGPT using natural language, rather than having to input specific commands, to make the experience more intuitive and seamless.
Thank you David. Your channel really opened my mind about AI's. I live in a poor place and I would never be able to pay for the content you teach here for free. Thanks
is the underpaid kenyan workers thing real?
thanks alot for awesome content. could you do video for fine-tuning customer support chat
Yed format is also text based and..you have a great visualiser
You only get about ten to 20 responses at a time though. It glitches and freezes before you get to finish anything
Probably lawyers adviced against including legal texts such as laws. They are smart people and thought deep about the consequences of having a model with knowledge of the law.
Thanks for the video tutorial. I have been on your github and installed a few! Thank you. So you have a link to the py.exe? I would like to be able to do tat myself! Cheers.
If you want to do knowledge graphs I believe generally the mainstream approach is to extract entities (people, places, things) from your documents, and then the graphing tool displays the relationships. There are open source "entity extraction" tools that vary in focus and in quality. (This has nothing to do with ChatGPT ) There are also graphing tools like Gephi.
This method isn't scalable. You still have to work within the token limit.
I see that6 the other programs that blend images is now a question of copyright, for the pool of images are owned by others. Fed District Court Nth Cal
insta sub, I know a wise sage when I see one.
Lots of great ideas, information, and how-to knowledge here, so thank you! I can't help but point out though that the title is very misleading, no? There's no training going on and we can't query chat GPT directly about all the documents as is implied. Or am I just missing something?
I too, have a lot going on. Yep! BIG things are in store for me friends. BIG things.
"frozen" : => "continue"
Chat GPT-3 for MCL concept:
Important question for any AI/IT/coding specialists:
I noticed this legal review regarding chat GPT and they trained it on supreme court case law.
I'm wondering if something similar could be retooled to be used with GPT-3 for the Michigan compilation of laws for Bill/policy research.
This could make legislative stuff and research far easier and less of a time consuming process.
(Would this be useful for Michigan law/current/past proposed laws and the parsing through all that info quickly for any research? In theory you could ask it specifically about any Michigan laws relating specifically to a specific topic, prior law, context, issue, parameter/quantity etc. So I could ask it about any laws relating to the legal amount of a specific type of chemical within a fertilizer in relation to current agricultural standards.)
Would anyone be interested in such a project/AI upgrade to current bill research?
This video deserves it's weight in gold.
https://youtu.be/E_sMa3N44u4?t=1491 you might be able to get the JSON-LD entities with IBM Watson which has an API for it.
Sometimes chatgpt would say that it doesn't know about something, but when I asked something related to it, it would go on and explain it with details. Like when I asked it about Hermes Warp Protocol, it said it knows nothing about it, but then it explained about it when it answered my question about Velo cryptocurrency.
"boxing matches sued Don King… oh this is fun!" David's reaction there is priceless (14:43)! Rewatched it like 3 times! Great video!
Thank you for your generosity to share! Its exactly what I have been searching and searching for!
Ohhh!!!! This explains why I asked for five court opinions in which the Wyoming court pierce the corporate veil and the defendant was a Wyoming LLC and it listed five cases in seconds with proper citations. However, when I verified the list, all of them were made up. I thought it could function as a basic search tool. Apparently, the origins of chatGPT that analogizes to brain nuerology means it likes to talk, create, … well, chat. I think my prompt was taken as a conversation starter. It tried to predict where the convo was going and court opinions were not facts, rather they were a way of expressing ideas. It is trained t avoid plagirism so it spoke in a predictive way using case law as language. I don't full get why it cant summarize a court opinion but it does not . I asked for an article on a specific IRS opinion letter and it could not get the fact pattern right. It was kicking out all kinds of weird facts
Thanks for the content, I saw all the 3 parts of the series but ultimately I did not get how you can make ChatGPT understand Case Law and pass it a corpus that GPT3.5 has not seen before. Did I miss something?
thanks for introducing me to the idea of using GPT to generate knowledge graphs from text, looking forward to exploring that in other contexts, the possibilities feel dizzyingly endless!
Subscribed. Needing info to prep for lawsuit.
this is really interesting!!! Thanks David.. Hoping we can go through this till the end
langchain can do this also, as i know.. it contains splitters, and have something written in it, so gpt was told to have memory for those splitted PDFs. though for now i haven't clearly checked the file of langchain..
for knowledge graph, there's a way to do it. a snippet in vs code called mark map can do it. I often ask gpt to respond in the form of markmap.and then i put the answer into markmap, here we go..
Great video! I’m an attorney. I am fascinated by what you are doing and how it can help people use AI to get legal questions answered.
Here are some thoughts that could be helpful to your goals with this project. To cram more useful text into ChatGPT without running into word limits, you could remove unnecessary words. For Supreme Court cases, unnecessary words include concurring and dissenting opinions, which are written opinions from justices that are not binding on subsequent cases. Every Supreme Court case starts with a majority decision (which becomes the law of the land) followed by a concurring decision, if any, and then a dissenting opinion, if any.
If you have legal questions, feel free to ask me. I would be happy to do what I can to contribute to your good work.