Abhijeet-LightweightFineTuning-Project

This is an example of using parameter efficient fine-tuning (LoRA) on a sequence classification model using a downstream dataset.

Low ranked adaptation (LoRA) reduces the number of parameters to train making it possible to run fine-tuning on commodity hardware with a smaller compute and memory footprint.

Check out: https://github.com/huggingface/peft

Prerequisutes

• Python 3 - https://www.python.org/downloads/
• Pycharm (optional) - https://www.jetbrains.com/pycharm/
• Pip (to install dependencies) - https://pypi.org/project/pip/

Dev environment with at least 16 GB RAM.

Background reading

• Hugging Face: https://huggingface.co/
• Transformers API: https://huggingface.co/docs/transformers/index
• Datasets API: https://huggingface.co/docs/datasets/index
• Evaluate API: https://huggingface.co/docs/evaluate/index
• PEFT API: https://huggingface.co/docs/peft/index
• Auto classes: https://huggingface.co/docs/transformers/en/model_doc/auto
• Trainer: https://huggingface.co/docs/transformers/en/main_classes/trainer

Foundation model

• GPT2: https://huggingface.co/openai-community/gpt2

Training dataset

• Dair-AI/emotion: https://huggingface.co/datasets/dair-ai/emotion

Install dependencies in your virtual environment using:

pip install -r requirements.txt

Checkout: https://stackoverflow.com/questions/41427500/creating-a-virtualenv-with-preinstalled-packages-as-in-requirements-txt

Run script

The script is included as main method. You can run the script using pycharm or via command line.

Current Limitation

Currently not using Quantization, since Quantization library (bitsandbytes) currently only has support for CUDA and lacks support for Metal (GPU available for Mac)