Dude, this video production quality is sick. Keep it up bro.
@jesseschultz6512
Жыл бұрын
I really do appreciate that you start with an over view and quick refresher before diving into the specifics. Managed to keep it short and to the point as well.
@youssefaserrar2001
Жыл бұрын
This is a well made video, great editing and explaining. Keep up the good work.
@jjphorn
Жыл бұрын
These are very much needed, great job! I agree with others, just slow down a bit and these will be perfect.
@vrsen
Жыл бұрын
Got it
@syntaxerror2872
Жыл бұрын
excellent work pal
@vrsen
Жыл бұрын
Thanks!
@vintagegenious
Жыл бұрын
straight to the point
@maxamad13
Жыл бұрын
Your video are really good man. Only downside is you go through them really fast. Wish you you explained the code little further. Appreciate your content.
@vrsen
Жыл бұрын
Thanks bro. Sorry, I will try to slow down on code next time
@2361244
Жыл бұрын
Few years ago Model fine tuning videos were more like a guy sharing his screen and walking me through the libraries used and why. Now it's a different world......kinda miss those in depth tutorials
@deepakkrishna837
10 ай бұрын
Hi Mate, great video. Can you please recommend which layer to be fine tuned if I want to build a custom document parser which can extract exact information from a document text when prompted with a question, like what is the education of a person if the document is his resume?
@deepakkrishna837
10 ай бұрын
Hi VRSEN can you please reply to my query?
@radouane5591
Жыл бұрын
good video, have you done a detailed video on finetuning MPT7B on a specific dataset?
@vrsen
Жыл бұрын
No, what are you building?
@radouane5591
Жыл бұрын
@@vrsen thank you for your reply. I want to finetune mpt7b on reviews dataset.
@etopowertwon
Жыл бұрын
Can you go even deeper? Like from renting runpod/vast ai instance(with enough of storage) and finetuning custom text dataset (just download some dataset like TinyStories, but pretend it's custom to show how it can be preprocessed) With occasional copying of intermediate weights to onedrive/mega/or simply scp-ing it to local machine (crucial step - I don't trust cloud further than I can throw them) Google's Colab Pro is not available everywhere. And problem with paid clouds it's really nerve-wracking to figure out what packages need to be installed as you are on the timer. I think I've spent two hours on cheapest offering from my GPU provider (single GTX1080) simply to find my way around 40GB of ubuntu lts1804 (so much stuff to delete)
@bhaveshchoudhary3189
Жыл бұрын
Please provide the solution. It will really help.
@vrsen
Жыл бұрын
Yes for sure. The next video is on QLoRA, which can help you save a lot of memory when fine tuning btw
@vintagegenious
Жыл бұрын
Yes please :)
@vivekmathur3068
Жыл бұрын
I wish to Fine Last 3 layers of MPT-7B model with the Data within a CSV and then within a PDF.. How should I do this? I want to fine tune it , and then ask it Questions interactively.
@vrsen
Жыл бұрын
Split it into inputs and outputs. Use code interpreter to ask gpt4 how to format your data and use it in transformers
@LinaKasmi-wy4zv
Жыл бұрын
That was really helpful thank you so much! Can you please make a video about how to fine tune pre-trained models like MPT-7B to build a chatbot ??
@vrsen
Жыл бұрын
You can use the same code and just change the prompt template for chat models
@LinaKasmi-wy4zv
Жыл бұрын
@@vrsen thank you🥰 I am new to AI world so I would like to ask what are the best pre-trained models for conversational text generation that can be used when building a chatbot for customer support?
@AndrewOrekhoff
Жыл бұрын
Thanks 🙏👍 exactly what I need
@vrsen
Жыл бұрын
Awesome, Andrew. What else would you like to see?
@AndrewOrekhoff
Жыл бұрын
@@vrsen Deployment and creation API of fine tuned model
@vrsen
Жыл бұрын
@@AndrewOrekhoff Great! Will do!
@TrelisResearch
Жыл бұрын
I was trying to adjust the collab code to use a longer sequence length. Any tips on configuring that. This approach doesn't seem to have worked... # Load the configuration config = transformers.AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote_code) # Explicitly set the max_seq_len config.max_seq_len = 4096 # Load the model with the updated configuration self.model = transformers.AutoModelForCausalLM.from_pretrained( model_name, config=config, cache_dir=cache_dir, torch_dtype=torch_dtype, trust_remote_code=trust_remote_code, use_auth_token=use_auth_token, ) thanks for having made the vid, my collab kept on crashing when I used simpler HF code.
@vrsen
Жыл бұрын
Hmm, what model are you trying to load?
@TrelisResearch
Жыл бұрын
@@vrsen mpt-7b-instruct It all worked fine with your code, it's just can't get the config change to work for max_seq_len. The hugging face repo recommends this: import transformers name = 'mosaicml/mpt-7b' config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True) config.max_seq_len = 4096 # (input + output) tokens can now be up to 4096 model = transformers.AutoModelForCausalLM.from_pretrained( name, config=config, trust_remote_code=True )
@TrelisResearch
Жыл бұрын
@vrsen any further tips here to configure the max_seq_len ?
@farziverse4209
Жыл бұрын
How is it diff from Qlora?
@vrsen
Жыл бұрын
It doesn’t add any additional parameters (adapters) to the model. You train only the last layer instead. I would recommend qlora for more general fine tuning on larger datasets, and this for more specific tasks.
@yasindukamizuru2377
Жыл бұрын
Hi i have a question so if i want to fine tune a question answering model using MPT i only have to change the dataset right?
@vrsen
Жыл бұрын
Yes, if your dataset is in the right format. Possibly adjust the prompt template as well
@vintagegenious
Жыл бұрын
Do you know how I can finetune MPT-7B to produce json data? It's always the same json, so I may use MS Guidance or Jsonformer, but how would I do the finetuning with the json dataset ?
@vrsen
Жыл бұрын
Yes, your dataset should contain ideal json outputs. It should be pretty large though. You can also add a prompt like “output json” to all inputs in your training set. This should help a bit.
@vintagegenious
Жыл бұрын
@@vrsen Do you think 10.000 is enough?
@nickynicky782
Жыл бұрын
Very good video, just Label? How would an inference with longer texts such as abstracts, articles, questions and answers be?
@vrsen
Жыл бұрын
You can basically reuse the same code, however, the model might not pick up very complex patters with this technique for long text generation. My next video takes a look at QLoRA, which should perform better for this use case. We will be fine tuning falcon 7b for question answering on free google collab
@nickynicky782
Жыл бұрын
@@vrsen Thank you very much for the answer, I have tried to fine tune with QLora with longer text but an error appears in the training, I would really appreciate it if you could explain how it could be done. thank you.
@balajimani2950
Жыл бұрын
I'm getting error with dataset and more over how to use multiple gpu in this method
@balajimani2950
Жыл бұрын
@vrsen
@vrsen
Жыл бұрын
Multiple gpus should work the same way. What is the dataset error?
@balajimani2950
Жыл бұрын
@@vrsen pls share the code for multigpu
@anllelomatamorosfruttero9184
Жыл бұрын
Very good video, you have ig?
@vrsen
Жыл бұрын
Nah, maybe later
@ankitjadhav3344
Жыл бұрын
Your videos are great sir. RESPECT 🫡
@user-ew8ld1cy4d
Жыл бұрын
This channel is FANTASTIC! Great ideas on content! PLEASE reconsider your next video on Falcon with QLORA. @venelin_valkov has already done 2 phenomenal videos on that. What a lot of us are looking at in the HF community is QLORA on your above video: MTP-7B. It seems a little bit more tricky... Just a suggestion. Either way awesome channel, keep it up and thank you!
@vrsen
Жыл бұрын
Thank you for suggestions! Already recorded QLORA with falcon, however I will be showing instruct model, and the code can be easily reused with MPT. Let me know what else you’d like to see
@AndrewOrekhoff
Жыл бұрын
Please send me your email, can't find it in channel info
Пікірлер: 55