My AskAI
Create an AI customer support assistant for your SaaS
Start new thread

No Code AI Model Builder - Create custom GPT-3 AI models in minutes, with No Code

by•
Build a custom GPT-3 model from csv data today in 3 clicks - no coding required. Create a WhatsApp or Telegram chatbot of yourself, a customer service responder based on your own chat history or something to write your own tweets for you. Let's start building!

Add a comment

Replies

Best
Maker
👋 Howdy PHers! With all the buzz around chatGPT and GPT-3 recently Alex and I have been coming up with tons of ideas of things we can build with them (same?), but most ideas need a bit more than the 'generic' GPT-3 models. The real powerful use cases are when you 'fine-tune' GPT-3 on more specific data sets. The problem I had though (and as a result, Alex had...) is that I am technically pretty inept and so had to ask Alex to train the model each time. We decided rather than keep doing this we'd just build a tool to make it a couple of clicks: 1. Choose the data you want to fine-tune the model with (in csv format) 2. Upload and submit it (we'll validate and optimise for OpenAI best practices) 3. Get back a fine-tuned model in a few minutes You can then use your model in the OpenAI playground, via API or through Zapier (for the full no-code experience!) Can't wait to see what people start to use it for, a few ideas we had: - Export a WhatsApp/Telegram chat history with a friend/relative and create an auto-responder based on you - Export your Intercom/Crisp support chats and create a personalised support chatbot - Upload your last few thousand tweets and get it to automatically generate new ones based on your tweeting - Export your previous blogs and get it to write new posts in your own voice But tbh we are looking forward to seeing what other people come up with! Try it out for free 👉 And pleaaaaase give us feedback, we'd want to improve it ASAP!
Congratulations on this big step / launch. Really rooting for your success.
Maker
Thanks Rhymer :)
It is not working at the customisation stage. The page just loads when I select the option to train the model.
Maker
ok no worries, are the headers of the two columns formatted saying exactly (same case): prompt completion ?
yes, I was using the bad jokes training data to test it out
Congrats on the launch! Been keeping track of this one for a while. A question though, what's the minimum amount of data I need to train a model? You mention I can create a chatbot of myself, but collecting the data for that would be a pain if I had to scrape together a lot of it.
Maker
so for generative models OpenAI recommends >500 prompts approximately. If you use WhatsApp or Telegram and do a chat export with someone you talk to regularly I imagine you’ll hit those kinds of numbers easily!
Thanks! How long should the prompts ideally be?
Maker
Sorry but that's a little bit of a 'how long is a piece of string' type question! Generally I'd say the more detailed and specific the prompt the better the output. In total prompts plus responses can't exceed 2,048 tokens (approx 8,000 characters I think) so worth bearing that in mind.
Oh, I understand now. I thought you were referring to fine tuning, but you really meant prompt tuning (fine tuning can definitely be over 2048 tokens, in fact, more is recommended haha). Very interesting though! Do you plan to incorporate other models in the future? I ask because I'm currently working on an API for an open source model (BLOOM) on the cloud, and I think it would be useful if more people had access to these models as a GPT-3 alternative.
Maker
yeah so our goal is essentially to make AI more accessible to people, so if there are other models out there that we could give simple access to then we’d be all for that! Happy to chat more about it if you want to contact me on Twitter
Hey Mike, seems like your Sendgrid account hit capacity. Wasn't able to register.
Maker
noooooooooo let me check!
Maker
(thanks for the heads up Vincent!
Maker
Hmmm I just tried and it seemed to work, would you mind trying again and maybe recording your screen with Loom so I can see what is going on? (send it to me on Twitter if you don't mind!)
sendspark(dot)com/share/xj2g3niezpijt0mszx2kdw2tss2vwb7c
Maker
thanks so much! Will get this sorted ASAP and let you know
Wow definitely gonna try this.
Maker
great! What are you going to do with it?
Thinking about testing a custom GPT-3 model with some awareness of the writing style of individual users. This is for a note-taking app my team has been building. Have you experimented with use cases like this?
Maker
oh interesting! I have not, we made a model to write personal statements for university students and played around with some cover letter writing! I can see it working well for the use case you speak about though if you get enough examples of their writing!
Yeah thanks for sharing. It sounds interesting, and I'll give it a try. Another quick question - is this tool capable of continuous import of data with incremental retraining of the model? Or is there any plan to support this in the future? For my use case, it might be critical for the AI to learn and improve itself in certain way.
Maker
Currently no, but can certainly add this as something for us to consider in the future if we have enough demand for it!
Wow! Thanks for making this. I am going to try it.
Maker
Thanks Zeng! Looking forward to hearing what you think of it! Did you have a use case in mind?
Awesome work Mike, I will train it with my bad jokes!
Maker
Hahaha thanks Peter!
Maker
please do!
Congrats! This looks great 🎉
Maker
Thanks Stefan!
Thank you for building such a helpful app! 🚀 I am a fan of this app!
Maker
no problemo Vitalie! Glad you like it!
this looks incredible! Is it possible to input data with over +1500 words per prompt? Also how many prompt do you recommend to create a good model?
Maker
hey! So each prompt and completion can’t be more than 2,048 tokens according to OpenAI’s docs, and I believe 1 token is approx 4 words, so will let you do the maths on that one! Generally OpenAI recommends for a good generative model you should have 500 or more examples to begin with. Does that answer your questions?
Very cool idea guys!
Maker
Thanks Serdar, hopefully we can start making fine-tuning your own models more mainstream!
123
Next
Last