t5-gqa-moe-xsum_with_lb

Model Description

T5 with Grouped Query Attention (GQA) + MoE fine-tuned on the XSUM dataset for abstractive summarization.

Architecture

This model combines two advanced techniques:

1. Grouped Query Attention (GQA): Custom attention mechanism that reduces KV cache size
2. Sparse Mixture of Experts (MoE): Multiple expert networks with learned routing

Key Features:

• Custom GQA implementation (not library-based)
• Token Choice Top-k routing with load balancing
• Reduced inference memory footprint

Model Configuration

• Base Model: google-t5/t5-small
• Total Parameters: 229,656,576
• Trainable Parameters: 229,656,576
• Number of Experts: 8
• Top-k: 2
• Load Balancing: Enabled
• GQA Query Heads: 8
• GQA KV Heads: 2

Training Data

The model was trained on the XSUM dataset, which contains:

• ~204k training examples
• ~11k validation examples
• ~11k test examples

Each example consists of a BBC news article and a one-sentence summary.

Usage

from transformers import T5Tokenizer

# Load tokenizer
tokenizer = T5Tokenizer.from_pretrained("YOUR_USERNAME/t5-gqa-moe-xsum_with_lb")

# Note: For MoE models, you need to reconstruct the architecture
# See the model repository for detailed loading instructions

Evaluation

Evaluate using standard ROUGE metrics and SummaC consistency scores.

Training Procedure

The model was trained using:

• AdamW optimizer with weight decay
• Learning rate: 5e-5
• Warmup steps: 500
• Mixed precision (FP16) training
• Gradient accumulation for larger effective batch size

Limitations

• Trained only on English news articles
• May not generalize well to other domains
• MoE models require custom loading code

Citation

If you use this model, please cite the XSUM dataset:

@inproceedings{narayan-etal-2018-dont,
    title = "Don{'}t Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization",
    author = "Narayan, Shashi and Cohen, Shay B. and Lapata, Mirella",
    booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
    year = "2018",
}