JohnDoe4765/MoE-Transformer-Model-Zoos

MoE Transformer Model Zoos

The MoE Transformer Model Zoos consist of two datasets: AGNews-MoEs and MNIST-MoEs. These datasets contain small-scale Mixture-of-Experts (MoE) Transformer weights and are designed to support research on metanetworks for predicting model generalization directly from internal weight structures.

Folder Structure

The dataset is provided as a data.zip file. After extraction, the directory structure is organized as follows:

data/
├── ag_news/
│   ├── weights/          # Directory containing model checkpoints
│   └── ag_news.csv       # Metadata file (accuracies and hyperparameters)
└── mnist/
    ├── weights/          # Directory containing model checkpoints
    └── mnist.csv         # Metadata file (accuracies and hyperparameters)

Dataset Overview

Dataset	Task	Samples (Checkpoints)	Description
MNIST-MoEs	Image Classification	100,024	Pretrained weights on the MNIST handwritten digit dataset.
AGNews-MoEs	Text Classification	79,220	Pretrained weights on the AG News topic classification dataset.

Each entry includes weight checkpoints at epochs 50, 75, 100, and the epoch of peak accuracy, alongside training and test accuracy records.

Model Architecture

Both datasets utilize a consistent Transformer-MoE architecture:

• Blocks: 2 Transformer-MoE blocks.
• Experts: 4 experts per MoE block, each consisting of a two-layer feedforward network.
• Classifier: Global average pooling followed by a two-layer MLP with ReLU activation.
• MNIST Specific: 2D convolutional embedding with fixed positional encodings.
• AGNews Specific: Word2Vec token embeddings with fixed positional encodings.

Hyperparameter Configurations

Models were trained by exhaustively combining hyperparameters across two configuration families:

Hyperparameter	SGD / SGDm	Adam / RMSprop
Top-K	[1, 2, 4]	[1, 2, 4]
Activation	ReLU, GeLU	ReLU, GeLU
Train Fraction	[1.0, 0.9, 0.8]	[1.0, 0.9, 0.8]
Dropout	[0.2, 0.15, 0.1, 0.05, 0]	[0.2, 0.15, 0.1, 0.05, 0]
Init Std Dev	[0.1, 0.15, 0.2, 0.25]	[0.1, 0.2, 0.3, 0.4]
Learning Rate (MNIST)	[1e-3, 3e-3, 5e-3, 1e-2, 3e-2]	[3e-4, 5e-4, 1e-3, 5e-3, 3e-2]
Learning Rate (AGNews)	[1e-3, 3e-3, 1e-2, 5e-2, 7e-2]	[3e-4, 1e-3, 5e-3, 3e-2, 5e-2]
L2 Reg (MNIST)	[1e-6, 1e-4, 1e-2]	[1e-6, 1e-4, 1e-2]
L2 Reg (AGNews)	[1e-8, 1e-6, 1e-4]	[1e-8, 1e-6, 1e-4]

Computing Resources

The datasets were generated using a cluster of 4x NVIDIA A100 SXM4 80GB GPUs.

• MNIST-MoE training time: 20–25 minutes per setting.
• AGNews-MoE training time: 30–35 minutes per setting.

Purpose

These datasets facilitate the study of weight-space modeling, specifically for MoE architectures, allowing researchers to train metanetworks that predict performance without accessing original test data.