PEFT-U: Parameter-Efficient Fine-Tuning for User Personalization at Scale

This repository consists of the code and data for the paper Parameter-Efficient Fine-Tuning for User Personalization at Scale. In this work we explore the problem of personalizing LLMs to support user centered tasks where the preference of users' can potentially differ for the same input. Conventional LLMs focus on having one model to support all users, we emprically analyze the use of prompting LLMs vs tuning and compartmentalizing user-level knowlege for personalized tasks.

Python version 3.10.9.

Personalized Training

Installation

For PEFT methods, e.g. LoRA, P-tuning and personalized head, run

pip install -r requirements_peft.txt

For Adapter methods, run

pip install -r requirements_adapter.txt

Two separate environments are needed since adapter-transformers directly modifies HuggingFace’s transformers and overrides their identifiers.

Add current directory for python to look for our local package:

export PYTHONPATH=$PATHONPATH:`pwd`

All scripts are intended to be run from the root directory.

Process datasets

Run files in peft_u/write_data.

For example, process the TweetEval dataset, run

python peft_u/write_data/prepare_gabhate.py

Run Training

PEFT

Arguments

Shared arguments for train and test sub-parsers

• model: HuggingFace model name or local model path
• dataset_name: Dataset name
• leakage: whether to allow text sample leakage between train and test set across users
  • Current experiments are w/ leakage == True
• method: PEFT frameworks, one of [lora, prefix, p_tuning, prompt_tuning]
• seed: random seed for 1) loading dataset demo examples and 2) training
• batch_size: batch size for training/evaluation

Additional arguments for the train sub-parser

• num_epochs: Number of training epochs
• learning_rate: Optimizer learning rate
• weight_decay: Optimizer weight decay
• output_dir: Output directory name postfix
  • Note an output directory name will be generated from the training arguments

Additional arguments for the test sub-parser

• zeroshot: If given, a single model is loaded for inference on all users with no personalization
  • Intended for evaluating zero-shot performance

Example

1> Run Prefix Tuning training with the EPIC dataset:

python peft_u/trainer/baseline_peft.py train --dataset_name 'epic' --method 'prefix'

2> Run inference on a a model trained with UnhealthyConversations

python peft_u/trainer/baseline_peft.py test --dataset_name 'unhealthyconversations' --model '23-06-03_{md_nm=flan-t5-base, ds=unhealthyconversations, peft=p_tuning}'

Adapter

Arguments

The same as that of PEFT, except that

1. method: Adapter methods, one of [Houlsby, IA3]
2. There’s no zeroshot flag for the test sub-parser

Example

1> Run IA3 training with the TweetEval dataset for 16 epochs, with output postfix 16ep

python peft_u/trainer/baseline_adapter.py train --dataset_name 'tweeteval' --method 'IA3' --num_epochs 16 --output_dir '16ep'

2> Run inference on a model trained with the WikiDetox dataset

python peft_u/trainer/baseline_adapter.py test --dataset_name 'wikidetox' --model '23-06-22_{adapter=Houlsby, md_nm=flan-t5-base, ds=wikidetox}'

Personalized Head

Arguments

The same as that of PEFT, except that there’s no method argument, for the only method is Personalized Head.

Example

1> Run Personalized Head training with the StudEmo dataset:

python peft_u/trainer/personalized_head.py train --dataset_name 'studemo'

2> Run inference on a model trained with the TweetEval dataset

python peft_u/trainer/personalized_head.py test --dataset_name "tweeteval" --model "23-08-05_{PH, md_nm=flan-t5-base, ds=tweeteval}"