You explained very well, but I want to know how do I run the model present on original tatsu_lab-stanford-alpaca on my laptop? I cant figure out how to do it... everywhere there are other repositories like dalai and etc I dont want to use those.
Yes. In my new video (to be released today) I show you how to set the target modules mapping to your specific LLM, plus if you want to apply parameter-efficient fine-tuning (with LoRA), if your GPU has not enough memory for fine-tuning the bigger LLM versions (like 11B+) .
Thanks, great Video and great explanation! But doesnt finetuning a model (with Data of ChatGPT) for commercial purposes break the Terms&Conditions of OpenAI?
Therefore when you sign the contract to use the products of OPENAI, like Stanford did and payed for the product sold by OPENAI, there is the fine-print w/ all the license restrictions and for a university I think, you can negotiate favorable terms. But I do not have access to their contract, therefore I can't answer your legal problem. I would ask OPENAI directly if you need further info or special conditions for your corporation.
In theory, of course. it is just a huge array of tensor multiplication. However, since companies create their model with their own special tokens, their own little secrets, you have to know about their configuration approach ...
i was thinking the same thing, imagine using a model like gpt for the main model that can interact with 10 alpaca models for its internal processing for different tasks. could be the steps to a primitive agi :P
An ALPACA model is just a fine-tuned model for one downstream task on a LLama model from Meta. Therefore an average general intelligence (AGI) seems theoretically possible. However, adding the different learned weights from all layers in the LLM together, catastrophic forgetting is happening, like in 2019. The solution was a fusion adapter for all fine-tuned adapters, with frozen main weights.
It should be noted that this 52k training data was generate with text-davinci-003 which was the previous best model for this task.. but now OpenAI has released gpt3.5-turbo which is better, faster and most importantly 10x cheaper.. so if you wanted to repeat this exercise it would probably cost a lot less than $500 USD. (I doubt the university discount was 10% of the commercial price). The biggest problem with this method is it may violate OpenAIs Terms of Service so use at your own risk.
Correct me if I'm wrong, but if you don't plan to compete with OpenAI then you're not violating terms when you train a model...? Otherwise surely they'd explicitly say you cannot train any non-OpenAI models... what they seem to be saying is don't attempt your own commercial LLM
@@runitup7029 yes so you can only get away with it for educational/research purposes. the moment you try to commercialize any product based on that model you step into a legal minefield because it probably violates both Meta and OpenAIs current TOS. but then again.. OpenAI and Meta have already stepped into a legal minefield by "training" their models on copyrighted materials just like Stability AI and Midjourney have with their image generation models.
this is the best video on fine tuning alpaca yet.
Glad it was helpful!
Wow, this was a great dive into the model. These guys really thought outside of the box. Impressive.
I know, right?
I think your next video should be on making custom datasets using other models like chat GPT etc
You explained very well, but I want to know how do I run the model present on original tatsu_lab-stanford-alpaca on my laptop? I cant figure out how to do it... everywhere there are other repositories like dalai and etc I dont want to use those.
By run I mean Build, Train and Predict
Got clone the repo. Any internet search will show you how to do it on your machine (either PC, any Cloud platform, ...)
@@code4AI yes i cloned the repo. i dont want to generate data so i dont need open ai API.
@@code4AI Do i need to get LLAMA model from somewhere else? it is not included in alpaca's repo
one more query, i am unable to run because i have intel iris gpu
Hum... it's kind of cool 12:31
Thanks for the vidéo !
You are welcome!
Amazing work !!!!
Wait so can we add more data sets to our liking? Can you train Alpaca with GPT 3.5?
Yes. In my new video (to be released today) I show you how to set the target modules mapping to your specific LLM, plus if you want to apply parameter-efficient fine-tuning (with LoRA), if your GPU has not enough memory for fine-tuning the bigger LLM versions (like 11B+) .
@@code4AI sweet can’t wait
Thanks, great Video and great explanation! But doesnt finetuning a model (with Data of ChatGPT) for commercial purposes break the Terms&Conditions of OpenAI?
Therefore when you sign the contract to use the products of OPENAI, like Stanford did and payed for the product sold by OPENAI, there is the fine-print w/ all the license restrictions and for a university I think, you can negotiate favorable terms. But I do not have access to their contract, therefore I can't answer your legal problem. I would ask OPENAI directly if you need further info or special conditions for your corporation.
Can such models collaborate with each other and work like a team in on a specific task/ app ?
In theory, of course. it is just a huge array of tensor multiplication. However, since companies create their model with their own special tokens, their own little secrets, you have to know about their configuration approach ...
i was thinking the same thing, imagine using a model like gpt for the main model that can interact with 10 alpaca models for its internal processing for different tasks. could be the steps to a primitive agi :P
An ALPACA model is just a fine-tuned model for one downstream task on a LLama model from Meta. Therefore an average general intelligence (AGI) seems theoretically possible. However, adding the different learned weights from all layers in the LLM together, catastrophic forgetting is happening, like in 2019. The solution was a fusion adapter for all fine-tuned adapters, with frozen main weights.
ISO has to be given access to this technology and open source standards should be established!
How do I make a custom dataset?
Here is the file: github.com/tatsu-lab/stanford_alpaca/blob/main/generate_instruction.py
hi，I am interested in knowing whether this model can be trained to simulate a personalized girlfriend. Is this feasible? 0-0
While this potentially possible, I wouldn't recommend it.
This is amazing to use. Too bad it is still for research purposes only
But we are allowed to learn every little bit we can from this system ...
It should be noted that this 52k training data was generate with text-davinci-003 which was the previous best model for this task.. but now OpenAI has released gpt3.5-turbo which is better, faster and most importantly 10x cheaper.. so if you wanted to repeat this exercise it would probably cost a lot less than $500 USD. (I doubt the university discount was 10% of the commercial price).
The biggest problem with this method is it may violate OpenAIs Terms of Service so use at your own risk.
Correct me if I'm wrong, but if you don't plan to compete with OpenAI then you're not violating terms when you train a model...? Otherwise surely they'd explicitly say you cannot train any non-OpenAI models... what they seem to be saying is don't attempt your own commercial LLM
@@runitup7029 yes so you can only get away with it for educational/research purposes. the moment you try to commercialize any product based on that model you step into a legal minefield because it probably violates both Meta and OpenAIs current TOS. but then again.. OpenAI and Meta have already stepped into a legal minefield by "training" their models on copyrighted materials just like Stability AI and Midjourney have with their image generation models.