Fine Tune Palm 2 + How to make a dataset

[@paulocoronado2376]

Great video, Sam!! You are awesome!! I’ve got some cool ideas for future videos: 1. Advanced examples of PaLM2 + Langchain (expression language, custom tools etc.) 2. Langsmith (very useful for managing and troubleshooting models)

[@samwitteveenai]

Certainly will make some Lang Smith vids and some more with Expression Languages

[@Pure_Science_and_Technology]

Thank you Sam for putting this video together..

[@fontenbleau]

Also they would never allow downloading because Palm2 can be used to operate robots by camera and many more as people say on LLM Bootcamp convention.

[@sergeyscorin]

Hi. Cool video. Could you please tell me is it possible to set a book as a training data to learn it and than ask things from this book to get instant general answers from it ?

[@권소예대학원생-컴퓨]

Hi. Thank you for your work. Where can I get the 'reduced_output.json' file?

[@EstebanAstudillo]

Thank you Sam, very useful as usual! I'm wondering how much money did you ended up spending while using the vertex services? Or did you have a trial? (I missed that). Thanks!

[@eddiehaug]

Regarding the number of training steps. According to the docs: Train steps: Enter the number of steps to run for model tuning. The batch size varies by tuning location: * us-central1 has a batch size of 8 * europe-west4 has a batch size of 24 If there are 240 examples in a training dataset, in europe-west4, it takes 240 / 24 = 10 steps to process the entire dataset once. In us-central1, it takes 240 / 8 = 30 steps to process the entire dataset once.

[@sebastiansanchez7331]

Hello Sam, Is this supposed to take a while once I click "start tuning" button? I don't know if it's just me but once I click start tuning, it doesn't show me anything nor looks it's training.

[@oxytic]

Great tutorial thank you for doing this , kindly create autonomous angent tutorial using LLM please

[@ahmadzaimhilmi]

Fine-tuned openai models cost more to run inferences than the base ones. How about this?

[@samwitteveenai]

Currently from everything I have seen Google isn't charging more for the FT models, This may change though.

[@sakshatkatyarmal2303]

Awesome video..Do you know how to make a dataset for chat-bison for finetuning on gcp?

[@MuaazOsaid]

Hi Sam thank u for the video. im facing this issue Invalid file extension. Supported extension: 'jsonl' on vertex-pipelines-prompt-validation node, any help will be appreciated.

[@samwitteveenai]

I think they have changed this since I made the video. Something similar for Gemini is coming so I will make some new vids for that

[@pranaymehta7958]

I followed the steps for emoticon fine tuning and it is over 20 minutes that the step 12 / 19 is still running. The job is large-language-model-tuner. I wonder when this will be over. The dataset size is so small and the time taken to tune the model is so high.

[@pranaymehta7958]

finally finished after 1 hour. thank you for the video , this was super helpful and I got a good idea of the internal workings. Lets see how much costly this one comes out to be.

[@dedekj26]

are we can just use the fine tuning feature from google colab?

[@samwitteveenai]

Not for these models. We can FT the open source models in Colab. We can call the APIs to FT from Colab though.

[@waneyvin]

Great video Sam, Is there any chance to have a tutorial about RLHF as well?

[@samwitteveenai]

currently putting one together

[@waneyvin]

cheers mate!@@samwitteveenai

[@sam_joshua_s]

I love ur video and can you make video about fine tuning using deepspeed

[@mohsenghafari7652]

hi. please help me. how to create custom model from many pdfs in Persian language? tank you.

[@samwitteveenai]

Hi the keys to this will be 1. getting the data out of PDF (check out unstructured.io for this) 2. fine tuning on a model that has a good tokenizer. Gemini is probably a good fit for this.

[@mohsenghafari7652]

@@samwitteveenai tank you. link unavailable !

[@mohsenghafari7652]

@@samwitteveenai dear ! do you have a linke for a sample? tanks

[@tommhk]

i got "The following quota metrics exceed quota limits:" no matter what Accelerator type i chose (i have upgraded to pay account)

[@samwitteveenai]

it is probably in relation to quota of GPU or TPU in that region. Go to the quota section and apply for quota in those regions

[@grym6857]

not me already using me 300$ free trials while palm was subpar :(

[@kishantripathi4521]

Unable to fine tune with free tier

[@iskrabesamrtna]

@@kishantripathi4521 ofc not :)

[@888felipe]

Never try this. Just to train this emoji example cost me $40. The KZbin video is great, but he never mentioned the cost... Maybe because he wanted the views. I could just send $40 to him instead...

[@samwitteveenai]

wow that seems high. When I made the video they were still in the testing phase. I am genuinely surprised it cost that much. I will check it out and put a warning in the description.