Dataset Card for the LoRA WiSE benchmark

The LoRA Weight Size Evaluation (LoRA-WiSE) is a comprehensive benchmark specifically designed to evaluate LoRA dataset size recovery methods for generative models LoRA-WiSE spans various dataset sizes, backbones, ranks, and personalization sets, as presented in the "Dataset Size Recovery from LoRA Weights" paper.

Task Details
Dataset Description
Dataset Structure
- Data Subsets
- Data Fields
Dataset Creation
Citation Information
🌐 Homepage: https://vision.huji.ac.il/dsire/
🧑‍💻 Repository: https://github.com/MoSalama98/dsire
📃 Paper: https://arxiv.org/abs/2406.19395
✉️ Point of Contact: mohammad.salama3@mail.huji.ac.il

Task Details

Dataset Size Recovery Setting: We introduce the task dataset size recovery, that aims to determine the number of samples used to train a model, directly from its weights. The setting for the task is as follows:

The user has access to n different LoRA fine-tuned models, each annotated with its dataset size.
It is assumed that all n models originated from the same source model and were trained with identical parameters.
Using only these n observed models, the goal is to predict the dataset size for new models that are trained under the same parameters.

Our method, DSiRe, addresses this task, focusing particularly on the important special case of recovering the number of images used to fine-tune a model, where fine-tuning was performed via LoRA. DSiRe demonstrates high accuracy in this context, achieving reliable results with just 5 models per dataset size category.

Dataset Description

We present the LoRA Weight Size Evaluation (LoRA-WiSE) benchmark. More specifically, it features the weights of 2050 Stable Diffusion models, which were fine-tuned by a standard, popular protocol of dreambooth via LoRA. Our benchmark includes stable diffusion version 1.5 and version 2, having 1750 and 300 fine-tuned models for each version respectively. We fine-tune the models using three different ranges of dataset size:

Low data range: 1-6 images.
Medium data range: 1-50 images.
High data range: 1-1000 images.

For each range, we use a discrete set of fine-tuning dataset sizes. In the low and medium ranges, we also provide other versions of these benchmarks with different LoRA ranks and backbones. See Data Subsets for the precise benchmark details.

Dataset Structure

The dataset contains seven subsets, each comprising 250-300 LoRA fine-tuned models. Each row in the dataset represents a single fine-tuned model, containing all the necessary information for recovery and numerical evaluation.

Specifically, each sample's dataset row corresponds to a single fine-tuned model with 256 layers, adding two new columns: "label" and "name." The "label" indicates the number of samples used for the dataset size of the fine-tuned models, while the "name" denotes the name of the micro-dataset.

We decided to provide the LoRA layers' weights (adaptive weights) instead of the full model for two reasons:

Providing the LoRA weights significantly reduces the storage size of the dataset.
Offering the LoRA weights enables users to study the properties of the fine-tuned LoRA layers, which may aid in developing new methods.

Data Subsets

The table below describes the dataset subsets in detail:

Subset Name	Dataset Sizes (labels)	Source	Backbone	# Of Models	# LoRA Rank
high_32	[1, 10, 100, 500, 1000]	ImageNet	SD 1.5	250	32
medium_32_2	[1, 10, 20, 30, 40, 50]	ImageNet	SD 2	300	32
medium_32	[1, 10, 20, 30, 40, 50]	ImageNet	SD 1.5	300	32
medium_16	[1, 10, 20, 30, 40, 50]	ImageNet	SD 1.5	300	16
low_32	[1, 2, 3, 4, 5, 6]	Concepts101	SD 1.5	300	32
low_16	[1, 2, 3, 4, 5, 6]	Concepts101	SD 1.5	300	16
low_8	[1, 2, 3, 4, 5, 6]	Concepts101	SD 1.5	300	8

Data Fields

As described above, each row of the dataset represents a single fine-tuned model that should be recovered and contains the following fields:

name - The name of the micro-dataset that the model was fine-tuned on.
label - the number of images used for the fine-tuned model
lora_{lora_name}_A_weight - The LoRA A weight matrix of the LoRA fine-tuned models layer.
lora_{lora_name}_B_weight - The LoRA B weight matrix of the LoRA fine-tuned models layer.

where {lora_name} is the name of the layer of the LoRA fine-tuned model in the subset.

Dataset Creation

The fine-tuning of the the models was performed using the PEFT library on Concept101 and ImageNet datasets.

For the full list of models and hyper-parameters see the appendix of the "Dataset Size Recovery from LoRA Weights" paper.

Citation Information

If you use this dataset in your work please cite the following paper:

BibTeX:

@article{salama2024dataset,
  title={Dataset Size Recovery from LoRA Weights},
  author={Salama, Mohammad and Kahana, Jonathan and Horwitz, Eliahu and Hoshen, Yedid},
  journal={arXiv preprint arXiv:2406.19395},
  year={2024}
}