Skip to content

Exploration on-device self-supervised collaborative fine-tuning of large language models with limited local data availability, using Low-Rank Adaptation (LoRA). We introduce three distinct trust-weighted gradient aggregation schemes: weight similarity-based, prediction similarity-based and validation performance-based.

License

Notifications You must be signed in to change notification settings

epfml/personalized-collaborative-llms

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Personalized collaborative fine-tuning for LLMs

This is the code base for the paper Personalized Collaborative Fine-Tuning for On-Device Large Language Models

Quickstart

Install conda environment:

conda env create -f env.yml && conda activate llm-lora

Generate dataset manually

If you want to generate heterogeneously distributed datasets before running fine-tuning:

python ./src/gen_dataset.py <dataset_name>

Base command

This is base command to run suitable experiments for Nvidia A100 GPU:

python ./src/main.py --config_format lora --use_pretrained gpt2 \
--eval_freq 25 --pretraining_rounds 100 --iterations 500 \
--lora_mlp --lora_causal_self_attention --lora_freeze_all_non_lora \
--trust <trust> --dataset=<dataset_name> --num_clients <num_clients>

Reproducibility

To reproduce the experiments you can run the following commands. There are several point to take note of:

  1. This might take a while to run, because we run the experiments with 10 different seeds.
  2. There two variants to the datasets.
  3. For clearer visualisation on wandb, we used a different project for each dataset.
  4. The number of clients used can be changed but there is a specific range for each dataset (maximum 20 clients and must be a multiple of the number of categories it has).
  5. Trust-based collaboration weight to aggregate the models:
    • "none": only local training
    • "naive": using collaboration weights 1/N
    • "static": using theoretical fixed weights as described in the paper
    • "dynamic": our strategy 1, based on model similarity
    • "dynamic-ref": our strategy 2, based on validation performance
    • "dynamic-token": our strategy 3, based on prediction similarity

General script

To get the same results, you can just run the following command, it will create a wandb project per dataset and have 10 runs per pair (trust, dataset).

./scripts/reproduce.sh

AG News

News articles in four categories: “World”, “Sports”, “Business” and “Sci/Tech”.
dataset_name = <"agnews_mixed"|"agnews_specific"> num_clients = <4|8>

./scripts/script.sh <wandb_project> <dataset_name> <trust> <num_clients> 10

Multilingual Wikipedia

Wikipedia texts in three languages (categories): French, Italian, and German. dataset_name = <"three_multi_mixed"|"three_multi_specific">

./scripts/script.sh <wandb_project> <dataset_name> <trust> 9 10

Codes-Wikipedia (Eng)

The first category is Java code from GitHub (HuggingFace) and the second category is English Wikipedia text dataset_name = <"github_wiki_mixed"|"github_wiki_specific">

./scripts/script.sh <wandb_project> <dataset_name> <trust> 8 10

Code structure

src                              # Main source folder                                                                                                                                                                                                                                                               
├── config                         # Config files for differents models
│   ├── __init__.py                # Chooses the correct config
│   └── lora.py                    # Configuration parameters for LoRA models
├── data                          # Datasets folder
│   ├── agnews.py                  # AGNews dataset
│   ├── fed_cc_news.py             # Comon Crawl News Federated Leraning dataset 
│   ├── github_wiki.py             # Java github code and wikipedia EN dataset
│   ├── three_multi.py             # Wikipedia FR, DE and IT
│   ├── utils.py                   # Get the correct dataset from the configuration parameters
│   ├── wikitext.py                # EN Wikipedia, single client test
│   └── wikitext_split.py          # EN, FR, IT or DE, single client test
├── distributed                   # Distributed package to run experiments on multiple GPU, NOT used in our experiments
│   └── ...                        # Only the default, single backend is used
├── gen_dataset.py                 # Allows to generate only datasets
├── main.py                        # Put everything togheter to fine-tune LoRA model on the various dataset
├── models                        # Models definition
│   ├── lora.py                    # LoRA nano-GPT model definition
│   └── utils.py                   # Get the correct model from the configuration parameters
└── optim                         # Training/Fine-tuning loop for the models 
    ├── lora.py                    # Loop for LoRA model
    ├── strategies.py              # Aggregation methods for collaborative learning
    └── utils.py                   # General useful methods

About

Exploration on-device self-supervised collaborative fine-tuning of large language models with limited local data availability, using Low-Rank Adaptation (LoRA). We introduce three distinct trust-weighted gradient aggregation schemes: weight similarity-based, prediction similarity-based and validation performance-based.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •