imbalanced-llm's Introduction

Imbalanced-LLM

Classification of Imbalanced Data with LLM (Large Language Model)

This project code uses TabLLM and T-few project and its paper.

As an Undergraduate Researcher, I did not had the luxury to have over 30GB GPU, and did not want to spend money on Colab. So I had to modify lots of code and versionings to make it work on free tier Colab, and even locally with smaller LLMs.

If there are questions, please leave an issue on github to talk more about it, or email me at [email protected]

Version

This code is ran on Google Colab Free Tier. It will follow those versions.
Used Python Black Formatter

Folders

/.old: Old attempt for Imbalanced LLM. Testing Idea.
/bin: Shell code to run the project
/configs: Configuration Data, related to /src/utils/Config.py
/Datasets: Raw csv datasets (Not included, go to TabLLM Project)
/Datasets-serialized: Serialized datasets (Not included, go to TabLLM Project)
/exp_out: Your Train result (Not included)
/pretrained_checkpoints: Saved Model (Not included)
- If using model T0 or T0_3b, get the file from TabLLM Project, turn on load_model() in EncoderDecoder() and add file
/src: Your Source
/templates:

Testing

This tutorial is modified from TabLLM Project.

We will use Google Colab Free Tier to run.

How to add your own Dataset

For my case, we will use stroke-prediction-dataset/data

1. Serialize Dataset

Run Make_Datset.ipynb Or
create_external_datasets.py --dataset stroke

2. Add files

Go to evaluate_external_dataset and add your dataset name on args_datasets variable
Make a new file called template_<datasetName> on templates folder. Use other templates as reference.
Go to bin/few-shot-pretrained-100k.sh and add your dataset on for dataset in <dataSetName>

Train/Fine Tune

Run TabLLM.ipynb

Get Result

For Stroke Dataset

python src/scripts/get_result_table.py -e t5_\* -d stroke

imbalanced-llm's People

Contributors

Watchers

imbalanced-llm's Issues

Train with setfit

Use setfit provided by hugging face and see if good to use for a few shots.
I will tweak some code to make it beneficial for our research of imbalanced data.