Code for Bayesian low-rank adaptation for large language models

This library is largely based on Laplace and ASDL.

Before installing laplace-lora, first install ASDL from source

pip install git+https://github.com/kazukiosawa/asdl

To install laplace-lora, change directory to laplace-lora and run

pip install -e.

To fine-tune LlaMA2 or any GPT-like model on common sense reasoning tasks, use

accelerate launch run_gpt.py

or the bash file

bash run_gpt.sh

for submission to a slurm server. Customize training arguments like lora_alpha, lora_r, lora_dropout, etc. Set testing_set argument to val if using the full training set; set testing_set argument to train_val to split the training set into training and validation set.

There are several hyperparameters that can be tuned for LoRA fine-tuning, e.g. lora_alpha, lora_r, lora_dropout, learning_rate, etc.

To use the full training set and Laplace model evidence for optimizing Laplace prior precision, set the testing_set argument to val; to split training set into a training set and a validation set and use minibatch gradient descent on the validation negative log-likelihood for optimizing Laplace prior precision, set the testing_set argument to train_val.

To run post-hoc Laplace approximation on saved checkpoints, use

accelerate launch run_gpt_laplace.py

or the bash file

bash run_gpt_laplace.sh

for submission to a slurm server.

To use full Laplace-LoRA, set the laplace_sub argument to all; to use last-layer Laplace-LoRA, set the laplace_sub argument to last_layer.