databricks-academy / large-language-models Goto Github PK

In LLM 03, we call models that have now been deprecated by OpanAI so that the commands fail when we run them.

This is the case in the demo where we define a call to text-babbage-001 which OpenAI says should now be replaced with gpt-3.5-turbo-instruct.

Further down in the Agents section, we use

from langchain.llms import OpenAI
llm = OpenAI()

which defaults to calling text-davinci-003 which is also deprecated and should be replaced by gpt-3.5-turbo-instruct.

Hey Team,

It will be great if you can provided the ipynb version.

Thanks,
Subham

Hi

Really comprehensive contents.
Can you please share if we can run all notebooks offline / without Internet connection after initial model download?

And is there any API access needed to run notebooks in this course?

Thank you

I get the below error, when i run the code in Jupyter Notebook:

Adding the below line of code solved the problem:

pip install -U accelerate and restart the notebook

Here is the error discussion: https://github.com/huggingface/transformers/issues/23340

Hello!

I could not find the "Releases" section in this repo for getting the slides of the course.
Can you please point out where can I find them?
Thanks,
Florin

This lab uses the CharacterTextSplitter which splits based on a separator not on chunk_size and chunk_overlap we need to either remove those or use the TokenTextSplitter

In LLM 03L - Building LLM Chains Lab notebook - Question 3, when the user tries map_rerank chain type, the model we use ('google/flan-t5-large') may not be powerful enough to consistently generate scores that can be parsed by the default parser in the default prompt in order for the results to be ranked.

The student may see error messages like

Could not parse output: [score between 0 and 100]

There are similar reports of this issue on langchain langchain-ai/langchain#3970

When I run

%run ../Includes/Classroom-Setup

for the first time, and I received errors. I looked inside the code, it was at the line "DA.reset_lesson()". I was able to bypass this by commenting out this line, since it was the first time running the code, there's no need to reset anything. But I think it may be helpful to provide the feedback.

In LLM 04 demo, we call imdb_ds = load_dataset("imdb") as our fine-tuning dataset.
It looks like there was an update to this dataset, and this line will throw an error ExpectedMoreSplits: {'unsupervised'}.

This can be fixed by forcing a re-install of the latest version of Hugging Face's datasets library. However doing so breaks the code further down where it can't find the train and validation splits in the dataset object.