Llm phi

[porteratzo]

A jupyter notebook to demo finetuning with the Phi3 LLM model. https://huggingface.co/docs/transformers/main/en/model_doc/phi3

We use the math_10k dataset https://github.com/AGI-Edgerunners/LLM-Adapters/tree/main

We download the dataset from github and check its hash and preprocess the dataset.

We load the phi model and prepare it for PEFT training.

Then we create a workflow that extracts the peft parameters and excludes sending the full model to the collaborators, the collaborators reconstruct the model from the original model and only modify the peft parameters. And after training and evaluation return the peft parameters again to be used with fed average.