Create Custom AI Characters Easily 🎭 How To Fine-Tune LLMs For AI Role Play
Вставка
- Опубліковано 2 жов 2024
- Thanks to Gradient for sponsoring this video: grdt.ai/mberman
In this video, I show you how to fine-tune an open-source LLM to behave like any character from movies, TV shows, comics...even yourself from text messages!
Enjoy :)
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewber...
Need AI Consulting? ✅
forwardfuture.ai/
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew...
USE CODE "MatthewBerman" for 50% discount
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Links:
colab.research...
Should I clone myself with AI and try to fool my family?
I would definitely love to see that!
Yes, check if mum is catching on
Hi @ mathew_berman , In your autogen courses, you've never used open interpreter. I wonder why
@@marilynlucas5128 I have an entire video of open interpreter :)
Yep.
I see a problem here: how do we know, that model responses "kinda" like Rick not because model knew of this character before finetuning? Any model can do what you did in this video right now, without finetuning.
What about an open-source tutorial. NO THIRD PARTY poo poos (I mean gradient is great and all, but come on!)
He does not seem to understand that using a product is not what "learning" means 🙁
Finally! I can automate my friendships, have the AI give me the highlights/details of any important information, and get back to living as a hermit.
🤣
Nice! Thanks Matt! Looking forward to this one. Introducing the real, the one, the only, Skippy The Magnificent!
Your videos have risen in quality and sophistication. Brilliant.
Thank you
Yes! Today I was researching how to fine tune Mistral 7B on my dataset and now you post this video! ❤
Would be great to also get a tutorial how to fine tune with HuggingFace AutoTrain. This way it could be uploaded to HuggingFace and used from Text-Generation-Webui.
Very cool! This is really another step into creating our own content... I love the last line where he is going to use the portal gun to go back in time and stop the portal gun from being invented! hilarious.
Could you please Do more videos on autogen specifically how to get autogen running with mem GPT and llm and maybe what are you using for if you're using it at all
Know what I'd really like to see? A guide on an open-source installation of multiple roleplay chatbots that can exist on a Discord server as Discord bots with the ability to communicate with each other and humans. That's my dream chatbot project, my server would love it.
As a weaponized autistic, the world is going to regret this tutorial.
You win the internet today i'm fkn dead, LOL
THANK YOU!!!
I was just getting ready to go down this exact rabbit hole when you posted this.
I WOULD LOVE to see a video of this BUT but using all open source/local llms/software.
I support this notion
This should be very easy!
I want to build a youtube social media audience, should I start my first video with that? Or a blog would be better?
Can you clone yourself and ask the model questions that you have answered before and compare the accuracy?
Two problems with the video:
1. You haven't compared the fine-tuned model with the original. Nous-hermes can roleplay at about this level with character prompt and no fine tuning so we don't really know if you did anything at all by fine-tuning the model.
2. The process for doing this locally for free has significant differences to what you shown, so it's not really useful as a tutorial. You wouldn't want to use collab for local training and you probably would use different libraries. Only data collection and preprocessing would be similar.
Can you share how you got to rick & morty wiki from Fandom? I search on rick & morty and got a ton of stuff that wasn't you showed in the video. Would also love to see a how to use Fandom for finding dialog from other movies & tv shows. I spent about 5 minutes (not a lot, I know) and really didn't find anything I thought I could use for fine tuning.
👎 You have to contact them for pricing.. Not my cup of tea.
How much would it have cost to train your example if it wasn't sponsored.❓
Can you do an fully open source free version of the same task? ❓
Thanks. Not really sure about any use case implementations. I think you mentioned using this anywhere, but there’s no context. Maybe you’re assuming most subscribers already have this knowledge?
Noob question: how to get the fine-tuned model?
Does GradientAI allow downloading the model/lora? or do we need to use its API to interact with the model?
Good question
Love your channel man! It's always the first thing I watch when it pops up. I have been playing around with famous characters using character sheets and Oobabooga. I've found that with only a few hundred tokens I can create a scary good interaction because (model dependent) everything is already baked in and all you need to do is point out who you want to talk to. I'm willing to bet that Rick and Morty script was part of your models training data set already. That said, AS ALWAYS, your video has filled my head with day dreams. The idea of training a whole model for a single character is so cool. My challenge to you and anyone that reads this is how small of a model can you train into being a coherent and usable roll play bot.
…You had me at Rick Sanchez. Subbed. 👍🏽
Amazing thank you!
1. How much data did you use there? How much is needed?
2. Why did you choose that base model?
nous Hermes is an unlocked Llama2 13B model - it’s more performant and also a little spicier
I use this through CHATGPT custom instructions and voice chat while I make dinner or work on code. I have Rick Sanchez roast me on everything like a Jerry! 🤣
Trick them, I want to see how they'll react,, Create the clone of you
Imitation episode would be awesome
Hello can anyone help me create an Alan Watts bot
It is really good but can you give it his voice and run it on his own ecosystem some thing like a raspberry pi with mic speaker....
honestly dude i feel like you specifically know what i'm doing and are making videos to help me. you keep giving me exactly what i need. i need ai agents to set up ai agents for me. smh. great video again! i'm saving this one
00:00 🎭 You can create a custom AI roleplay bot for characters from TV shows, movies, comics, and even yourself using fine-tuning techniques.
00:27 📊 Data for fine-tuning can be obtained from sources like Kaggle for scripts, which are available in CSV format, or other websites for various dialogues.
01:50 📄 You can also fine-tune a model based on your own text messages, DM history, or telegram messages by converting them into CSV format.
03:02 📂 Google Colab is used to prepare data for fine-tuning, and Gradient AI makes fine-tuning models easy and efficient.
05:36 ⚙ Fine-tuning is performed using NAOUS Hermes 2, and you can adjust the fine-tuning process as needed.
07:01 🗨 The model's responses are generated by providing a query, and you can customize the system message and instructions for roleplay.
08:50 🤖 You can deploy the AI roleplay bot for any character, and the process is versatile, allowing you to mimic various personalities.
Yes!! Would love to see you cloning urself
yeah! we need self clone with AI video. Eagerly waiting for your video sir
How long did it take roughly to complete the fine tune?
Yes definitely interested in chatting with a Mathew AI .. just don't trick your family members! 😂
I agree on both accounts
That's exactly what i wanted to try out, thanks! :) I would love to see if you can trick your family
I used to do this with markov chains way back, and had an IRC channel full of different characters that would discuss social media posts. One plus of markov chains was they could learn, but couldn't get that big, I wonder if llama index can replace the need for training.
Could you do some kind of tutorial of how to make a realtime voice conversation chatbot? So kinda like this where you train it to be the character you want, then add a speech-to-text and text-to-speech chat? Maybe something where you can add your own custom voices.
This is interesting and clever but it's pointless asking questions about its plan about anything. There's no permanence. It's just learning words and phrases and then outputs those to questions asked. The AI doesn't have missions and goals, likes or dislikes or any moral's, at least none that are permanently instilled in it. Not yet anyway.
>fine-tune my AI on my own data
>find them super annoying and hate their guts
>"Oh god, is that what I sound like?"
Okay demonstration, though I would strongly prefer seeing local approach.
Did you ever do a video on unstructured data for finetuning in the end?
I would love to see a video of you cloning yourself. Has that been made already?
Great stuff. Make a version to do this locally.
if i finetune model in gradient and then later decide to no longer use gradient, do i iose access to those models?
Does this only work for character behavior and personality or would it work with information if you loaded a database of writings, like, could it think like a person based on a dataset like that, so like talking to Einstein for instance?
In principle, yes - this is roughly how chatGPT and stuff were built. However it takes a ridiculous amount of data and compute power to actually change some more fundamental capabilities, that's why truly models with improved coding, better reasoning, better math etc. come out at a much slower pace then models with role playing this and that. The foundational models were trained with just a gigantic corpus of natural language and therefore they are extremely good at writing and translating between languages, but not that much with other stuff like science and the behaviour of the physical world.
@@auriocus I have seen where they are trying to make something where you can talk historical figures, but I think it is more novel than actually functional. I am really waiting for a way to have a Jarvis, like, an avatar on my pc, a head, or body, doesn't matter, but I can talk to it and have it open programs and perform tasks within those programs like 2 people working together on a project, but with knowledge to help me through problems that I may have. I don't see how this is that far off, internet access, speech capabilities, both talking and listening, and a memory system... I think we are close, like a true AI assistant.
ive been trying all of your videos i cant get a single one of them to work AHHHHHHH
Here an idea for for anybody willing to do it:
Fine tune a GPT model to help users install/configure Autogen locally, connect to various LLMs (including local ones) and thereafter quickly generate autogen code of various structures/configurations/tools - I would pay to use that model.
Just unsuccesfully spent several hours trying to get local Autogen going.
I don't see anyone mentioning this so: Here is a case where if you don't know something, the AI will confidently tell you something wrong. Rick doesn't believe in time travel... He would never postulate using a portal gun to "go back in time", but it sure sounded good to everyone who doesn't know that about Rick. Anyone who knows, would know this was an imposter Rick.
Who wouldn't! I mean, strictly for yuks of course, **and that point has to be emphasized with a disclaimer.** The subject would have to be selected carefully (the more sense of humor they have the better; not too young or old). Maybe do multiple friends/relatives and select the best one. The subject would have to be debriefed afterward, possibly on video. But given all this I think it would be a great video.
This is fascinating - all of your videos that I've seen so far are! Thank you so much for having a good fine-tuning video. I am very curious about two things that I am not seeing covered much... (1) The topic you mentioned - creating good prompts.. particularly from large bodies of unrefined data. For example 'all the works of Shakespeare' or 'the Guttenberg project'. Especially where those works have both dialog and non-dialog 'things' that happen as part of the writing style. (2) How to work entirely locally, assuming you have the memory / GPU for it - such as if you are working with a large body of local data and you don't want to keep trying to move it into the cloud during training, or you just want to use your own resources.
Yes, trick your family.
But how to do it without gradient that's not real opensource
DeepSeek Coder 6.7b Instruct is the only Open Source LLM that I was able to get to generate SNAKE!!!!! It works so well! And I tried the 33b version on their chatbot webpage and it also did it even better. I hope you make a video soon!
if "Rick == row["name"]? Sir, Yoda conditions do not have a reason to exist
Can this tell me how to make concentrated dark matter?
From where the model takes the informations besides the input .csv? And why the model making up responses if it dont know the answers? For what the input csv is, if I can chat with thant model with only the system message that inform it who he is?
I love the idea of cloning oneself, but is tricking someone into thinking they’re talking to a human when they’re not an unethical version of the Turing’s experiment?
my question is, now if i did this, how i use it as an api to incorporate it in my apps?
There are going to be breaking charges in the world of AI for the next several years as we all figure out exactly how to best use this AI thing.
Literally no idea what you did in this video.
I want to have a chat interface. I mean like a What'sApp chat with voice. That Chainlit thingy is still hard for me.
I was waiting for you to ask how to pronounce PENALTY. 😂Great video again.
make the video to trick your family....
Rick would actually do that lmao... He would go back in time to prevent the portal gun from being invented. But he would know he can't
do it, i think thats the next step, life tuning on your personality and create an alter ego bot of yourself, that would be epic
Missed opportunity, you should have asked Rick "What is my purpose?" :)
Thanks Matt for another great video, is the data accessible. How did you copy data from fandom and added in the file?. I would really appreciate if you can share that information.
how to do it without gradient?
That's Rick Sanchez alright 🤣Cool stuff!
Can I fine tune in other language than english?
BTW… there is a way to fine-tune model to use other language than English?
slow down! i'm one week behind
What’s up with so many AI content creators just pushing sponsored content. “You can do this without using Gradient” and goes on showing just Gradient. Sorry but the quality has gone down and I am unsubscribing, sadly.
Trying to earn a living. Thanks for your time while you were subscribed.
@@matthew_berman to be honest, I am not against it and I hope you get more of it. But dedicating a whole video under a misleading title isn’t fair to us “fans” of the channel. And to make it worse, the video doesn’t mention any other options just Gradient. “How to create custom AI characters easily using Gradient” would be more honest and less clickbaity. Of course, you do you but I’m just giving an honest feedback from someone who used to be one of the first to watch new videos. Someone has to tell an emperor that they aren’t wearing clothes, right?
I wonder if it would be possible to do this with video/audio of myself...
I absolutely want you to finetune a model, clone your voice with elevenlabs and trick your family member in a phone call
Nice video @matthew_berman , Do you know how can we implement this on structural data with multiple csv files tables in it and star schema.
yes i would love to see a video of you cloning yourself and tricking your family. Great work as always boss. so awesome..
rick doesn't do time travel.
Can I use this to fine tune my model that is not for AI roleplay?
I love this! Do you know any way to do clone voices as well? Goal: Generate a Rick Sanchez Voice using some GPT-generated text
Okay what about this AutoGen + memGPT + open interpreter?
Cloning yourself I think would be really cool. Would it be possible that instead of csv 2 columns to use audio files? Might be able to feed it an audio memoir or autobiography and make it more real and nuanced
Id like to see if you could get a good enough model to trick people you know for sure lol
YES and if you can make it work off a local llm that would be great
so how much did that all cost, and how do you download the new model?
Thanks for the vid - curious about the cost for this one...
is is possible to take fine tuned model from gradient to runpod? or can i fine tune in runpod?
Bring on the clones ❤😂
Do it!
love your channel. please bring back the AI News videos!
Yes do the clone and video keep up the good work you do I enjoy all your videos 🎉
Clone yourself 🎉
The man himself is back. I been looking forward to your uploads you always got so much dope information
Make the vid!
amazing! looking forward to have another video for this subject !!💙
How does this handle foreign language?
Yeah! Do it!!! The clone-yourself video! 🙌
I'm RickGPT!
Cool see ya in the next one. Bravo!
Very well done! Looking forward to the future fine tuning videos
You're incredibly helpful :D
YES we want that video! :)