GudiVaraprasad / Robust-LLMs-for-OOD

This study investigates multiple strategies with fine-tuned and pre-trained language models to evaluate and improve their robustness to out-of-distribution (OOD) data.

README : Robust LLMs that work in OOD settings

This repository contains scripts for fine-tuning Inference from differennt public models like GPT-2 , LLAMA etc testing LLMs FineTuning : full fine-tuning, LoRA, and QLoRA adaptation methods. generating synthetic data

Setup

pip install transformers torch numpy datasets
pip install git+https://github.com/path/to/peft.git 
pip install langchain
pip install openai==0.28
pip install -U langchain-community  
pip install -U sentence-transformers    
 

Usage

The main script is designed to be used from the command line with several options for customization:

• --model_name: The name of the model you want to fine-tune or test.
• --finetune: Specify the fine-tuning method (lora, qlora, or full_finetune).
• --infer: Enable this flag to test the model directly without fine-tuning.
• --test_data: Path to the test data file (default: "test_dataset.txt").
• --prompt: Prompt string for testing the model in inference mode.
• --data: Choose the data source for fine-tuning: synthetic data or an external dataset.

set your together api key in the local environment using this command

export TOGETHER_API_KEY=your_api_key

set your wandb key in the local environment using this command

export WAND_KEY=your_wandb_key

Example Command:

#Fine Tune Commands
python -m src.model --dataset yelp --finetune finetune_roberta
python -m src.model --dataset amazon --finetune finetune_roberta  —batch_size 16

#Prompt methods
python -m src.model --prompt_type reformulation --model_name gemma_2b #for rewriting 
python -m src.model --prompt_type explanation  --model_name llama_8b_it # for evaluation
python -m src.model --prompt_type CoT

# Evaluate
python -m src.model —-evaluate t5_sentiment

python main.py   --infer gpt "Which is the hottest planet?"



#FLIPKART Data
python main.py --flipkart Clean # to clean the data and save it to the disk use only once 
#YELP DATA CLEAN
python main.py --yelp Clean 
#imdb DATA CLEAN
python main.py --imdb Clean 

#PLOT EMBADDINGS:
python main.py --plot_embaddings ag

#FIne tune roberta on Yelp dataset
python main.py --finetune='finetune_roberta' --dataset='yelp'

Fine-Tuning Methods

• Full Fine-Tuning: Updates all parameters of the model.
• LoRA: Low-Rank Adaptation focuses on updating only a small set of parameters, reducing the computational cost.
• QLoRA: A quantized version of LoRA that further reduces the model size and computational requirements.
• finetune-roberta: full finetuning roberta model.

Data Generation

The script can generate synthetic data if the --data option is set to synthetic. This data is used for fine-tuning when no external dataset is provided.